WO2024104648A1

WO2024104648A1 - Technical design of an analysis device for spectral analysis

Info

Publication number: WO2024104648A1
Application number: PCT/EP2023/077531
Authority: WO
Inventors: Alexander Freytag; Erik Rodner; Karsten Lindig; Anselm BRACHMANN
Original assignee: Carl Zeiss Spectroscopy Gmbh
Priority date: 2022-11-14
Filing date: 2023-10-05
Publication date: 2024-05-23
Also published as: DE102022130044A1

Abstract

The invention relates to a method for technically designing an analysis device for the spectral analysis of at least one component of a sample. The analysis device comprises a spectrometer and is configured to use a chemometric model, which is formed by a trained machine learning system, to determine a concentration of the component based on spectral measurement values of the sample that are recorded by the spectrometer. A plurality of steps are performed iteratively (06). These steps comprise applying (07) a current configuration of a representation of spectrometer hardware, as a result of which spectral training data (08) is obtained. The chemometric model is trained (09) using the spectral training data (08). The chemometric model formed by the trained machine learning system is used to determine a value of a suitability characteristic (11) for the current configuration of the representation of the spectrometer hardware. The current configuration of the representation of the spectrometer hardware is modified (13). The invention also relates to: a computer programme product; a trained machine learning system; and an analysis device.

Description

Technical design of an analysis device for spectral analysis

The present invention initially relates to a method for the technical design of an analysis device for the spectral analysis of at least one ingredient of a sample. The analysis device comprises a spectrometer for the spectral measurement of the sample. The analysis device is configured to use a chemometric model formed by a trained machine learning system to determine a concentration of at least one ingredient of the sample based on spectral measured values of the sample recorded with the spectrometer. The ingredient is, for example, a protein or starch. The ingredient is, for example, water, so that a moisture content can be determined by the analysis device. The invention further relates to a computer program product and a trained machine learning system as well as an analysis device for the spectral analysis of at least one ingredient of a sample.

DE 10 2020 116 094 A1 concerns a method for calibrating a plurality of identical spectrometers for ingredient analysis. Using a mathematical model of the identical spectrometers, a large number of error spectra are generated in order to improve a regression model.

WO 2021/198247 A1 shows a method for co-designing hardware and software for virtual staining of a tissue sample. The method comprises iteratively obtaining multiple sets of training imaging data with respect to one or more tissue samples. Each set of the multiple sets of training imaging data was acquired using a different image modality of a group of imaging modalities. Multiple reference images are acquired representing the tissue samples containing one or more chemical dyes. The multiple sets of training image data are processed in a machine learning logic.

US 11,062,481 B2 relates to a portable device for determining a state of one or more plants. The device comprises a digital color camera for recording a color image of the plants within a field of view and a light source for providing broadband illumination for the plants within the field of view. A processing unit serves to control the camera and the light source during recording of a first image of the plants while the light source illuminates the plants with the broadband illumination and during recording of a second image of the plants while the light source does not illuminate the plants.

DE 10 2021 105 869 Al shows a spectral sensor system with an array of optical sensors arranged on an integrated circuit, as well as with an interface between the plurality of optical sensors and a first processing device. A plurality of sets of optical filters are configured as a layer arranged on a plurality of optical sensors. Each set of optical filters contains a plurality of optical filters. Each optical filter is configured to transmit light in a different wavelength range. The

Spectral sensor system comprises a processing device containing an artificial neural network that configured to correct a spectral response generated by the plurality of optical sensors.

US 2021/0172800 Al relates to techniques for analyzing unknown sample compositions using a prediction model based on optical emission spectra. First emission spectra are received that correspond to a training sample comprising several pure elements of known concentrations. Based on this, a plurality of spectral ranges are determined that correspond to the plurality of pure elements of known concentrations. Features are determined that are associated with a maximum of the spectral range. A prediction model is trained to predict unknown concentrations of a plurality of components of an unknown sample based on an emission spectrum of the unknown sample.

EP 3 842 788 A1 relates to a spectral sensor for near-infrared spectroscopy, which is used to distinguish and/or recognize objects and/or materials. The spectral sensor is designed to work in a learning mode and in an operating mode. In the learning mode, measurements of intensity values of a first spectrum of wavelengths are carried out. In the operating mode, measurements of intensity values of a second spectrum of wavelengths are carried out. The wavelengths of the second spectrum are selected with the aid of a machine learning method.

US 10 , 020 , 900 B2 shows a system architecture for providing spectral information for one or more devices and for providing and using the spectral information in a device. A Spectral Information Server executes a server process that has a library of spectrum-related functions; including functions for retrieving spectral data from spectral data sources or for processing retrieved spectral data.

The above-mentioned state-of-the-art solutions for spectral analysis provide for a spectrometer and an evaluation of the spectral measured values recorded with the spectrometer. Both the spectrometer and the method for evaluating the spectral measured values must be highly tailored to the given analysis task in order to be able to produce accurate results. Such a spectral analysis is carried out, for example, to determine the ingredients of agricultural products and foodstuffs. A chemometric model is used to evaluate the spectral measured values, which is developed, for example, using machine learning. The spectral sensitivities that the spectrometer should have must be specified; in particular, how the average wavelengths and bandwidths of the spectrometer’s measuring channels should be dimensioned. Selecting a spectrometer that is not optimally configured impairs the evaluation by the chemometric model. For example, the spectrometer may have too few or unsuitable spectral sensitivities, which limits the determination of the ingredients. In other cases, the spectrometer may be over-dimensioned for the given analysis task, i.e. it may have spectral sensitivities that are not required for the analysis task. This makes the spectrometer too expensive. US 5,435,309 shows a method for determining one or more unknown values of at least one known property, such as the concentration of glucose in blood. The method uses a model based on a set of samples with known values of the known properties and a multivariate algorithm. Several wavelength subgroups are selected from an electromagnetic spectral range suitable for determining the property for use by an algorithm. The selection of the wavelength subgroups is intended to improve the suitability of the model. The selection process uses multivariate search methods that select both predictive and synergistic wavelengths within the wavelength range used. The suitability of the wavelength subgroups is determined by a suitability function. A genetic algorithm is used to generate one or more counting spectra. Several counting spectra are combined to generate a combined counting spectrum. The count spectrum is smoothed and a threshold count value is selected from the count spectrum to select those wavelength subgroups that are suitable for optimizing the suitability function.

DE 10 2008 002 355 A1 shows a method for calibrating a spectrometer for measuring spectral components of a substance in the context of chemometrics. A spectrometer device is used to measure spectral components of the substance to be examined. The spectrometer device is calibrated using a calibration unit. A total of spectral components X and associated substance concentrations of the substance to be examined are measured and stored as spectrometric measurement data in the form of a multidimensional coefficient vector. From this From the totality of the measured spectral components X, physically relevant spectral components X _rei for the respective measurement process are extracted using an automated iterative estimation procedure stored in the calibration unit. The measured spectrometric measurement data for the substance under investigation are calibrated on the basis of the extracted spectral components X _rei that are physically relevant for the substance under investigation.

The object of the present invention, based on the prior art, is to be able to adapt the technical design of an analysis device for the spectral analysis of at least one ingredient of a sample more precisely to the respective given analysis task, whereby the analysis device can be designed in a less complex manner.

The stated object is achieved by a method according to the appended claim 1, by a computer program product according to the appended independent claim 13, by a trained machine learning system according to the appended independent claim 14 and by an analysis device according to the appended independent claim 15.

The method according to the invention is used for the technical design of an analysis device for the spectral analysis of at least one ingredient of a sample. The analysis device to be technically designed should be optimally suited to a given analysis task. The concentration of at least one ingredient of the sample is determined by the spectral analysis based on spectral information measured on the sample. The sample comes from a material or a product which is to be examined with regard to its content by the analysis. The sample can, for example, taken from the material or product or the material or product is fed to the analysis device as a sample. The material or product is preferably an agricultural product, a foodstuff or a foodstuff. The agricultural product is preferably a harvested product. The harvested product is preferably a grain such as corn or wheat or rapeseed, sugar beet or soy. The harvested product can also preferably be fruit plants such as peppers, tomatoes, strawberries etc. or fruit trees such as almond trees. The sample can be formed by the harvested product as such, as in the case of cereal grains, for example. The sample can also be formed by a part of a plant which produces the harvested product, so that, for example, the leaf of a cereal plant forms the sample. The ingredient is preferably formed by water, a protein, an oil, sugar, salt, starch or crude fibre. The ingredient is preferably formed by a single chemical element, such as nitrogen, phosphorus, potassium, calcium, magnesium, boron, molybdenum, copper, manganese, zinc, iron, chlorine or sulphur. The appropriate supply of nitrogen is of particular relevance for the growth of every crop, as nitrogen is the most important element of chlorophyll and is therefore essential for optimal metabolism. In addition to nitrogen, other macronutrients are also relevant, such as phosphorus and potassium, which are often added to the soil together with nitrogen as so-called NPK fertilizers. Calcium and magnesium are also relevant, the latter forming the central element in the chlorophyll ring. In addition, depending on the type of plant, other nutrients are often required for optimal growth.

Growth and optimal yield are relevant, which however are often required in much smaller quantities and are therefore also referred to as micronutrients. Examples of these are boron, molybdenum, copper, manganese, zinc, iron, chlorine and sulphur. Measuring the concentration of water represents a moisture measurement. The ingredient can also be an undesirable component in the product or crop being examined, such as a pesticide or a fungicide.

The analysis device comprises a spectrometer for spectral measurement of the sample. The spectrometer can be an NIR spectrometer, a VIS/NIR spectrometer, a VIS spectrometer or a full-range spectrometer. The spectrometer can have a transmission, a transflection or a reflection structure. The spectrometer is preferably designed compactly as a spectrometer sensor.

The analysis device is further configured to use a chemometric model formed by a trained machine learning system to determine a concentration of at least one ingredient of the sample based on spectral measurement values of the sample recorded with the spectrometer. The spectrometer records spectral measurement values of the sample, which are determined by the ingredients of the sample. By applying the chemometric model, the concentration of the at least one ingredient to be analyzed is determined from the spectral measurement values. The result is a value for the concentration of the at least one ingredient of the sample. The chemometric model is formed by a trained machine learning system. Training takes place when the method according to the invention is carried out. The machine learning system preferably comprises an artificial neural network or a linear model. which is preferably formed by a partial least squares regression model.

By means of the method according to the invention, the spectrometer and the chemometric model formed by the trained machine learning system are technically designed or developed so that they form the objects of a technical development process, the result of which is the technically specified spectrometer and the trained machine learning system.

In one step of the method, an initial configuration of a representation of a hardware of the spectrometer is selected. This representation can in particular be formed by a model or by a physical test example of the hardware of the spectrometer. The configuration defines the spectral sensitivities of the spectrometer. The initial configuration represents a starting point for the technical development process, which is changed during the technical development process. The result of the technical development process is a final configuration of the representation of the hardware of the spectrometer, which serves as a technical specification for producing at least one spectrometer, which forms a component of the analysis device to be produced.

The initial configuration of the representation of the hardware of the spectrometer is preferably formed by a configuration of the representation of the hardware of a reference spectrometer. The reference spectrometer is a high-resolution spectrometer, which is assumed to have all the necessary components for the given analysis task. spectral information. The reference spectrometer is a gold standard for this.

The following is an iterative execution of several steps which are computer-implemented. These several steps are executed iteratively. In simple preferred embodiments, these several steps are repeated until a predetermined number of iterations is reached. In further preferred embodiments, these several steps are repeated until a value of a suitability parameter has at least reached a predetermined limit value, which, depending on the definition of the respective suitability parameter, includes exceeding or falling below it. The one suitability parameter or the several suitability parameters each describe the suitability of the respective current configuration of the representation of the hardware of the spectrometer including the chemometric model formed by the trained machine learning system for the given analysis task. The at least one suitability parameter is preferably formed by a quality measure or particularly preferably by an error measure or a loss function. If the respective suitability parameter is formed by a quality measure, its value increases when the suitability increases. If the respective suitability parameter is formed by an error measure or loss function, its value decreases when the suitability increases. The predetermined limit defines how large or how small the suitability parameter must at least be so that the current configuration of the representation of the hardware of the spectrometer is considered suitable for the given analysis task, so that the spectrometer can be manufactured according to the current configuration for the analysis device. A first of the suitability parameters is particularly preferably formed by an error measure or a loss function , while a second of the fitness parameters is formed by a metric for selecting from the configurations .

In one of the iterative steps, the current configuration of the representation of the spectrometer hardware is applied, whereby spectral training data is obtained. If this iterative step is carried out for the first time, the current configuration of the representation of the spectrometer hardware is formed by the initial configuration of the representation of the spectrometer hardware. During application, a spectral measurement of reference samples is carried out or a spectral measurement of reference samples is simulated using a model of the spectrometer in order to obtain the spectral training data. The spectral training data represents spectral measurement data that is obtained with the current configuration of the representation of the spectrometer hardware for the reference samples by actually carrying out the measurement or by simulating the measurement.

In a further iterative step, the chemometric model formed by a machine learning system is trained using the current spectral training data. This training is a machine learning process, whereby the machine learning system is trained and the trained machine learning system is produced as a result. The machine learning is carried out with the aim of ensuring that the chemometric model formed by the machine learning system uses the spectral training data to determine the concentration of at least one ingredient in the best possible way according to the given analysis task. A wide variety of machine learning methods can be used to train the chemometric model formed by the machine learning system. In preferred embodiments, the chemometric model formed by a machine learning system is trained at each iteration of the iteratively executed steps, starting in an untrained state of the machine learning system. This means that no information from previous iterations is used. In alternative preferred embodiments, the chemometric model formed by a machine learning system is trained at each iteration of the iteratively executed steps, each starting in a pre-trained state of the machine learning system. This means that information or training experience from previous iterations or from previous training of independent, previous model optimization tasks can be used. The pre-trained state of the machine learning system can, for example, result from training that was carried out on the basis of the same spectral training data but to determine the concentration of a different ingredient in the sample. Similarly, the pre-trained state of the machine learning system can, for example, result from a training that was carried out on the basis of different spectral training data but to determine a concentration of the same ingredient in the sample.

In a further iterative step, the chemometric model formed by the trained machine learning system is applied to determine a value of at least one suitability parameter. The chemometric model formed by the trained machine learning system is applied to the spectral Training data. By applying this, the concentration of the at least one ingredient is determined for the current spectral training data. By comparing it with reference ingredient concentration data, the value of the at least one suitability parameter can be determined. Depending on the embodiment, the value of the at least one suitability parameter can be determined immediately after applying the chemometric model individually in each iteration or after executing all iterations of the iteratively executed steps for all applied configurations of the representation of the spectrometer hardware.

In a further iteratively executed step, the current configuration of the representation of the hardware of the spectrometer is modified. This modification is carried out in order to obtain a different configuration of the representation of the hardware of the spectrometer, which is used in the next iteration. The iterative modification is carried out with the aim of finding a configuration that is best suited to the given analysis task. The configuration of the representation of the hardware of the spectrometer is preferably modified within a predefined framework. This framework is preferably formed by at least one interval within which a parameter of the spectrometer defined by the configuration is changed. In simple preferred embodiments, the iteratively executed step of modifying the current configuration of the representation of the hardware of the spectrometer and the further iteratively executed

Repeat steps until the modified

Configurations the predefined framework with a predefined density. This can be done, for example, by changing a parameter of the spectrometer defined by the configuration at each iteration with a certain step size starting at a lower limit of the interval until an upper limit of the interval is reached, which can also be done for several parameters of the spectrometer defined by the configuration. This represents a grid search to find the most suitable configuration. Alternatively, a parameter of the spectrometer defined by the configuration can be randomly changed at each iteration until a predefined budget for this is exhausted. The budget is defined by a number of iterations to be performed and/or by a time period for performing the iterations. At each iteration, a parameter of the spectrometer defined by the configuration is randomly changed so that after a number of iterations defined by the budget, the predefined frame is covered with a predefined density and/or until the time period defined by the budget has expired. Density is to be understood here as meaning that in a configuration area there is a total number of configurations in this configuration area, whereby the configurations do not necessarily have to be evenly distributed over the configuration area. The number of iterations is preferably at least 100 and more preferably at least 1,000. After carrying out the iterative steps, the configuration of the representation of the spectrometer hardware that has a value of the suitability parameter that corresponds to the best possible suitability is preferably selected, so that the most suitable of the configurations used is selected. Thus, after the iterations, a decision is made as to which of the applied configurations is the one which

The result of the technical design of the spectrometer hardware can be seen in the described

Grid search can be a configuration which was generated by modification, for example, in the first iterations carried out or also in the last iterations carried out. The suitability parameter can, for example, be defined such that it decreases with the suitability of the configuration of the representation of the hardware of the spectrometer, so that the one of the applied configurations of the representation of the hardware of the spectrometer is selected whose value of the suitability parameter is minimal. This does not, of course, have to be the configuration of the representation of the hardware of the spectrometer applied in the last iteration carried out. As already explained above, it is particularly preferred to use a second of the suitability parameters as a metric for selecting from the applied configurations. The first of the suitability parameters and the second of the suitability parameters are preferably determined using different parts of the spectral training data.

In preferred embodiments, the iteratively performed step of modifying the current configuration of the representation of the hardware of the spectrometer is carried out using at least one of the previously determined values of the at least one suitability parameter. The at least one value of the at least one suitability parameter is used to improve the current configuration of the representation of the hardware of the spectrometer so that it can be better suited to the given analysis task, so that the concentration of the at least one ingredient can be determined more precisely. The The modified configuration of the representation of the spectrometer hardware represents the current configuration of the representation of the spectrometer hardware for the next iteration. In this embodiment, the iterative steps are preferably repeated until the value of the suitability parameter has at least reached a predetermined limit value.

Insofar as, as described above, this suitability parameter is also used as a metric for selection, the configuration of the representation of the spectrometer hardware used in the last iteration thus has the best suitability of all configurations used and represents the result of the technical design of the spectrometer hardware. Insofar as, as described above, a second of the suitability parameters is used as a metric for selection, the configuration of the representation of the spectrometer hardware used thereby represents the result of the technical design of the spectrometer hardware.

As already explained above, the at least one suitability parameter is preferably formed by a quality measure or particularly preferably by an error measure or loss function. The metric for selecting from the applied configurations is preferably formed by this quality measure, this error measure or this loss function. Particularly preferably, this metric is formed by a second of the suitability parameters. In preferred embodiments, a first of the suitability parameters is determined for modifying the current configuration using a preferably predefined part of the spectral training data, while a second of the suitability parameters is determined for selecting from the applied configurations using a preferentially determined other part of the spectral training data .

Alternatively, the iterative steps are preferably repeated until a predefined budget is exhausted, which is defined by a number of iterations to be carried out and/or by a time period for carrying out the iterations. After the iterative steps have been carried out, the configuration of the spectrometer hardware representation used is preferably selected whose value of the relevant suitability parameter corresponds to the best possible suitability, so that the most suitable of the configurations used is selected.

Particularly preferably, the modification of the current configuration of the representation of the hardware of the spectrometer is carried out using the current value of the at least one suitability parameter.

In a further preferred embodiment, in each iteration a parameter of the spectrometer defined by the configuration is changed randomly and/or using at least one of the previously determined values of the at least one suitability parameter. This is thus a hybrid of the embodiments described above. After carrying out the iteratively performed steps, the one of the applied configurations of the representation of the spectrometer hardware is preferably selected whose value of the suitability parameter corresponds to the best possible suitability, so that the most suitable of the applied configurations is selected. This does not, of course, have to be the one used in the last iteration carried out. Configuration of the representation of the spectrometer hardware .

The method according to the invention is characterized in that the spectrometer and the chemometric model formed by the machine learning system are technically designed or developed together. The spectrometer and the chemometric model together form the objects of a single technical development process. The hardware forming the spectrometer and the software forming the chemometric model are developed together, which can be referred to as co-design of the hardware and software of the analysis device. This co-design represents a significant difference to the prior art, according to which the spectrometer is first technically designed individually, for which the available knowledge is used as best as possible. In the next step, the chemometric model is developed on the basis of the technically fully designed spectrometer, which is done, for example, using machine learning. In contrast, in the method according to the invention the spectrometer and the chemometric model are developed together, which ensures that the spectrometer and the chemometric model solve the given analysis task in the best possible synergy.

A particular advantage of the method according to the invention is that the designed analytical device with its components, the spectrometer and the chemometric model, is very precisely adapted to the given analytical task. This ensures that the designed analytical device can fulfill the given analytical task very precisely. On the other hand, the spectrometer can be designed much less complexly. as a reference spectrometer according to the gold standard. The spectrometer designed as a result of the process is optimized and, in particular, only has those measuring channels with the respective bandwidths and average wavelengths that are required for the given analysis task. Many analysis tasks require only a small number of measuring channels, so this number is significantly reduced compared to the number of measuring channels of the reference spectrometer according to the gold standard. This reduces the costs for manufacturing the hardware of the analysis device. A further advantage is that the analyses can be carried out more quickly with the analysis device due to the reduced number of measuring channels and require less energy. Accordingly, less data needs to be transferred.

The chemometric model formed by the machine learning system is preferably trained using supervised learning. Supervised learning is preferably based on regression and/or classification. Regression is preferably used for continuously changing values of concentrations of an ingredient. An example of this is the concentration of water to determine moisture. Classification is preferably used for graded values of concentrations of an ingredient. An example of this is the concentration of salt, which is classified as either sufficient or insufficient. The concentration of salt can also be classified into levels; for example as: < 1.2 g/kg; 1.2 g/kg; 1.4 g/kg; 1.6 g/kg; 1.8 g/kg; 2.0 g/kg and > 2.0 g/kg.

The regression is preferably carried out for the partial smallest

Squares, which is called Partial Least Squares Regression ( Partial Least Squares PLS ) . This regression is robust and suitable when only a few latent variables have to be taken into account .

In a further preferred embodiment, the machine learning system comprises an artificial neural network, which is preferably formed by a convolutional neural network (CNN). The following parameters are preferably used: ID convolutions, non-linear activations, ID pooling operations and/or layer normalizations. A final prediction of continuous values of concentrations of an ingredient can be made using regression. A final prediction of graded values of concentrations of an ingredient can be made using classification, for example carried out as a multi-class classification, for example by predicting a discrete class number or by predicting a multi-class probability vector. The training can, for example, be carried out in such a way that a multi-class permutation error for discrete decisions or a multi-class cross entropy for probability vectors across multiple classes is minimized. In another implementation, an ordinal character of the classes can be taken into account, which is done, for example, by minimizing the weighted Cohen-Kappa loss value. Convolutional neural networks are suitable when a large amount of training data is generated and complex relationships exist.

Modifying the current configuration of the representation of the spectrometer hardware preferably using the current value of at least one Suitability metric represents an optimization of the representation . For this optimization, one or more optimization techniques from the following group are preferably selected and used: Bayesian optimization, grid search , random search , gradient-free optimization such as the Nelder-Mead method, optimal experimental design, machine learning methods , greedy algorithms, evolutionary algorithms, biology-inspired optimization methods such as particle swarm optimization and FireFly optimization . Gradient-free optimization is particularly preferred.

In a first group of preferred embodiments, the representation of the hardware of the spectrometer is formed by a model of the hardware of the spectrometer. The method then preferably comprises a further step in which reference data are provided which are used to apply the model of the hardware of the spectrometer. The reference data comprise a series of reference spectral data to which at least one series of reference ingredient concentration data is assigned. The reference spectral data would ideally be recorded from a sample which contains the ingredient according to the reference ingredient concentration data. The iteratively carried out step of applying the respective current configuration of the model of the hardware of the spectrometer comprises a simulation of a spectral measurement of the sample with this respective current configuration of the model of the hardware of the spectrometer. During the simulations, the respective configuration of the spectrometer hardware model is applied to the reference spectral data, thereby obtaining spectral simulation measurements that form the spectral training data. Thus, by applying the model developed by the trained machine learning system From the chemometric model formed, simulation measured values of the concentrations of the ingredient are obtained, which are compared with the reference ingredient concentration data in order to determine the value of at least one suitability parameter in each case. The aim is for the simulation measured values of the concentrations of the ingredient to be as close as possible to the reference ingredient concentration data. A particular advantage of the first group of preferred embodiments is that the spectrometer does not have to be physically present for its iterative improvement, since the spectrometer is represented by the model and iteratively optimized.

If the reference data is not already available from other sources, the step of providing the reference data is preferably carried out by first providing reference samples for which the concentration values of the at least one ingredient are known, so that these values form the reference ingredient concentration data. For this purpose, the concentration values of the ingredient can be determined using a chemical analysis method. These values can also be determined using a non-chemical analysis method; for example using a spectral analysis device with a high degree of accuracy. The reference samples are also measured using a reference spectrometer to obtain the reference spectral data.

In the first group of preferred embodiments, the method preferably comprises further steps, which take into account that the process of manufacturing the hardware of the spectrometer is not ideal. In principle, the hardware of the spectrometer, which is designed according to a model the hardware of the spectrometer will not completely resemble this model, but there will be a deviation from the model. For example, the mean wavelengths of measurement channels of the manufactured hardware of the spectrometer may differ from the mean wavelengths of the measurement channels of the model of the hardware of the spectrometer. Therefore, in one step, a manufacturing-related hardware deviation is first determined between the model of the hardware of the spectrometer and a piece of spectrometer hardware manufactured according to the model of the hardware of the spectrometer. The manufacturing-related hardware deviation determined is taken into account in the iterative step of applying the current configuration of the representation of the hardware of the spectrometer, so that the spectral training data correspond more closely to the measurement data that can be acquired with the manufactured hardware of the spectrometer. However, the manufacturing-related hardware deviation determined is preferably not changed during the iterations.

In the first group of preferred embodiments, the method preferably comprises further steps through which changing measurement conditions are taken into account in order to obtain extended spectral training data, through which the spectral training data as a whole becomes more realistic. Such measurement conditions are, for example, the distance between the sample and the spectrometer and environmental conditions such as air humidity and temperature in the area of the sample and/or the spectrometer. The changing measurement conditions are taken into account in the iterative step of applying the current configuration of the model of the spectrometer hardware, so that the spectral training data is more accurate. correspond to the measurement data that can be recorded with the spectrometer hardware under realistic measurement conditions. The measurement conditions are varied to take into account that different realistic measurement conditions can occur. For example, the change in the amplitude of a measurement channel can be changed depending on the ambient temperature. This amplitude represents the sensitivity of the measurement channel.

In a second group of preferred embodiments, the representation of the hardware of the spectrometer is formed by a test example of the hardware of the spectrometer. Thus, for each of the iterations, the hardware of the spectrometer must be physically provided again in the form of the test example; namely according to the respective current configuration. The method then preferably comprises a step in which reference samples are provided. The values of the concentration of the ingredient are known for each of the reference samples, so that these values form reference ingredient concentration data. The iteratively carried out step of applying the respective current configuration of the test example comprises measurements of the reference samples with the respective current configuration of the test example of the hardware of the spectrometer, whereby spectral measured values are obtained which form the spectral training data. The measurements are thus physically carried out with the test specimen in the current configuration for the physically available reference samples. By iteratively applying the chemometric model formed by the trained machine learning system, measured values of the concentrations of the ingredient are obtained, which correspond to the reference ingredient concentration data. be compared in order to determine the value of at least one suitability parameter.

For each iteration, a new or adapted test sample of the hardware must be provided. This can be done by creating the test sample. The test sample from the previous iteration can be reused for this purpose. In a simple case, only the filters of the measurement channels of the test sample are modified in order to provide a new or modified test sample for the next iteration.

Preferably, the respective test specimen is also used to measure reference samples, whereby spectral measurement values are obtained. These spectral measurement values are used as spectral training data in further iterations of the iterative steps. This takes into account manufacturing tolerances that occur when manufacturing the spectrometer.

For all groups of the preferred embodiments, the respective configuration of the representation of the hardware of the spectrometer preferably defines at least one parameter of the spectrometer. The respective configuration of the representation of the hardware of the spectrometer preferably defines several of the parameters of the spectrometer. The parameter(s) preferably specify at least one measuring channel of the spectrometer; more preferably at least two measuring channels of the spectrometer and even more preferably at least 16 measuring channels. The parameter(s) are preferably each defined by a peak wavelength, a center wavelength, a mean wavelength, a bandwidth, a half-width, a curve shape parameter and/or a peak value of the at least one measuring channel is formed. The peak value is a peak height. The one measuring channel or the multiple measuring channels are preferably each determined by a spectral input filter. The spectral input filters each have a transmission range which is defined by one or more of the above-mentioned parameters. During the iterative step of modifying the current configuration of the representation of the spectrometer hardware, one or more of the above-mentioned parameters are changed. The optimization techniques specified above are preferably used for this. In addition or alternatively, a number of measuring channels are changed during the iterative step of modifying the current configuration of the representation of the spectrometer hardware. For example, the number of measuring channels can be changed from 16 to 8 or to 4 or vice versa.

The representation of the hardware of the spectrometer preferably defines at least one limitation which is defined by a minimum peak wavelength, by a maximum peak wavelength, by a minimum centroid wavelength, by a maximum centroid wavelength, by a minimum mean wavelength, by a maximum mean wavelength, by a minimum bandwidth, by a maximum bandwidth, by a minimum half-width, by a maximum half-width, by a minimum waveform parameter, by a maximum waveform parameter, by a minimum peak value, by a maximum peak value, by a minimum wavelength difference between a wavelength parameter of one of the measurement channels and a corresponding wavelength parameter of another of the measuring channels or by a maximum wavelength difference between a wavelength parameter of one of the measuring channels and a corresponding wavelength parameter of another of the measuring channels. The minimum or maximum wavelength difference is defined in particular between wavelength parameters of two adjacent measuring channels. The wavelength parameter in question is preferably formed by the peak wavelength or by the minimum or maximum half-width. The restrictions mentioned are preferably also defined multiple times in each case, for example restrictions on the minimum or maximum peak wavelength in several intervals. The restrictions mentioned can also relate to a group or to a portion of the measuring channels. For example, for a portion of one third of the measuring channels it can be specified that their peak wavelength is in the near infrared range. The restrictions mentioned each limit the possible change to the parameter in question during the iteratively carried out step of modifying the current configuration of the representation of the hardware of the spectrometer. This one restriction or However, these multiple restrictions are preferably not changed during the iterative step of modifying the current configuration of the representation of the spectrometer hardware. This one restriction or these multiple restrictions are preferably read from a test copy of the spectrometer hardware. The representation of the spectrometer hardware preferably includes several of these restrictions.

The representation of the hardware of the spectrometer preferably further defines at least one device parameter of the spectrometer, which is a signal noise, a bandwidth Noise, a half-width noise, a measurement deviation or a tolerance of the spectrometer. The measurement deviation can, for example, be temperature-dependent. The aforementioned bandwidth noise is a deviation of an achieved mean bandwidth from a predetermined value for the bandwidth. The one or more tolerances are caused in particular by the manufacture of the spectrometer hardware. The device parameter(s) are taken into account in the iterative step of applying the current configuration of the representation of the spectrometer hardware, thereby improving the optimization of the spectrometer and the chemometric model. This is preferably done by augmentation during training, for which the simulation of the spectrum is changed in at least one iteration, taking into account the at least one device parameter. The device parameters are preferably not changed in the iterative step of modifying the current configuration of the representation of the spectrometer hardware.

The computer program product according to the invention is used for the technical design or development of an analysis device for analyzing at least one ingredient of a sample. The computer program product comprises a computer-readable storage medium which has program instructions stored thereon. The program instructions can be executed by one or more computers or control units and cause the one or more computers or control units to carry out the method according to the invention or one of the described preferred embodiments of the method according to the invention. The program instructions include, among other things: Algorithms for machine learning; namely for training the chemometric model formed by a machine learning system with the spectral training data. The storage medium can be formed by an electronic medium, a magnetic medium, an optical medium, an electromagnetic medium, an infrared medium or a semiconductor medium, such as an SSD. The program instructions can be formed by machine-dependent or machine-independent instructions, microcode, firmware, status-defining data or any source code or object code, for example written in C++, Java or similar or in conventional procedural programming languages. Electronic circuits, such as programmable logic circuits, field-programmable gate arrays (FPGA) or programmable logic arrays (PLA), can also be formed which execute the program instructions. The resulting model of the spectrometer hardware and the resulting chemometric model formed by the machine learning system are preferably stored as program code in order to be available for later applications .

The trained machine learning system according to the invention forms a chemometric model for determining a concentration of at least one ingredient of a sample based on spectral measurement values of the sample recorded with a spectrometer. The machine learning system according to the invention was trained by the method according to the invention or by a preferred embodiment of the method according to the invention. It is available in the form of a program code.

The analysis device according to the invention is used for the spectral analysis of at least one ingredient of a sample. The analysis device comprises a spectrometer for spectral measurement of the sample. The analysis device is configured to use a chemometric model formed by a trained machine learning system to determine a concentration of at least one ingredient of the sample based on spectral measured values of the sample recorded with the spectrometer. The spectrometer and the chemometric model formed by the trained machine learning system emerged from the method according to the invention or one of the described preferred embodiments of the method according to the invention. The analysis device was therefore technically designed or developed using the method according to the invention. The optimization of the spectrometer and the training of the chemometric model formed by the machine learning system took place together by carrying out the method according to the invention. The analysis device preferably also has further features that are described in connection with the method according to the invention.

In preferred embodiments, the analysis device comprises a plurality of components which are mechanically independent of one another and can have their own housings. The spectrometer is preferably designed as a handheld device, which forms one of the mechanically independent components. Another of the mechanically independent components of the analysis device is preferably formed by a computing unit which is configured to use the chemometric model formed by a trained machine learning system. The spectrometer and the computing unit are preferably connected to one another via a wireless data connection. The computing unit can be a smartphone, for example. Further details and developments of the invention emerge from the following description of preferred embodiments of the invention, with reference to the drawing. They show:

Fig . 1 : a flow chart of a first preferred

Embodiment of a method according to the invention; and

Fig. 2: a flow chart of a second preferred embodiment of the method according to the invention.

Fig. 1 shows a flow chart of a first preferred embodiment of a method according to the invention. The method is used for the technical design of an analysis device (not shown) for the spectral analysis of at least one ingredient of a sample. The method results in the analysis device to be developed being optimized for a given analysis task. The analysis device comprises a spectrometer (not shown) for the spectral measurement of the sample. The

The analysis device is further configured to use a chemometric model to determine a concentration of at least one component of the sample based on spectral measurement values of the sample recorded with the spectrometer. According to the method, the spectrometer and the chemometric model formed by a machine learning system are jointly developed and optimized. In this first preferred embodiment, a model of the hardware of the spectrometer is used to develop and optimize the spectrometer.

In a preparatory step 01, a

Reference spectrometer provided, which is high resolution and represents a gold standard. The reference spectrometer is chosen so that it provides a larger bandwidth and a higher spectral resolution than are necessary for the given analysis task. The spectrometer to be developed will therefore have a reduced bandwidth and/or spectral resolution compared to the reference spectrometer.

In a further preparatory step 02, reference samples are provided which are selected according to the given analysis task. For the reference samples, values of concentrations of at least one ingredient of the reference sample are known. This is the at least one ingredient whose concentration is to be determined according to the given analysis task. These values known for the reference samples represent reference ingredient concentration data. In a further preparatory step 03, the reference samples 02 are spectrally measured with the reference spectrometer 01, so that in a next step 04 several reference spectral data are available which are later used to simulate spectral measurements.

The method comprises several iterative steps 06, through which a joint iterative optimization of the model of the hardware of the spectrometer and the chemometric model in the form of the machine learning system takes place.

In an iterative step 07, the spectral measurement of samples with the spectrometer is simulated. For this purpose, a model of the spectrometer hardware is used. The model describes the spectrometer hardware by a specification (not shown) in Form of parameters. The specification s (not shown) is preferably formed by a vector. A first exemplary definition of the vector is described below: A first value of the vector s (not shown) is formed by an average wavelength of a first measuring channel of the spectrometer. A second value of the vector s (not shown) is formed by a bandwidth of the first measuring channel of the spectrometer. A third value of the vector s (not shown) is formed by an average wavelength of a second measuring channel of the spectrometer. A fourth value of the vector s (not shown) is formed by a bandwidth of the second measuring channel of the spectrometer. This sequence of values is continued accordingly. A second exemplary definition of the vector is described below: A first value of the vector s (not shown) is formed by an average wavelength of a first measuring channel of the spectrometer. A second value of the vector s (not shown) is formed by a bandwidth of the first measuring channel of the spectrometer. A third value of the vector s (not shown) is formed by an amplitude of the first measuring channel of the spectrometer. A fourth value of the vector s (not shown) is formed by a mean wavelength of a second measuring channel of the spectrometer. A fifth value of the vector s (not shown) is formed by a bandwidth of the second measuring channel of the spectrometer. A sixth value of the vector s (not shown) is formed by an amplitude of the second measuring channel of the spectrometer. This sequence of values is continued accordingly. The simulation of the measurements for recording spectra with the current model of the spectrometer hardware is carried out by applying the current vector s (not shown) to the reference spectral data 04, whereby spectral training data are available in a next step 08. When simulating the measurements for recording spectra for the first time, an initial configuration of the model of the spectrometer hardware is assumed, which is represented by an initial vector So (not shown).

The spectral training data presented in step 08 are used in a step 09 to train the chemometric model formed by the machine learning system. After each execution of step 09, the machine learning system is trained and is used to predict the value of the concentration of at least one ingredient in the reference sample from the spectral training data. Deviations of the predicted values from the reference ingredient concentration data represent an error which characterizes the current model of the spectrometer hardware. An objective function is therefore defined which indicates the aforementioned deviations for the respective vector s (not shown). The aforementioned deviations result in the value of a suitability parameter in a step 11, which is checked for a condition in a step 12. This condition forms the termination condition for the iteratively executed steps 06. This preferably predefined stopping condition defines when the model of the spectrometer hardware and the chemometric model formed by the machine learning system are considered sufficiently optimized to perform the given analysis task.

If the test in step 12 shows that the suitability parameter does not yet fulfill the condition , then in a step 13 modified parameters for the model of the spectrometer hardware are determined in order to achieve the objective function described above and thus also the suitability parameter . improve. Accordingly, in a t-th iteration 06 a vector s _t (not shown) is determined, for which preferably not only the immediately previously determined vector s _t -i (not shown) but all previously determined vectors So to s _t -i (not shown) are taken into account.

The determination of modified parameters for the model of the hardware of the spectrometer in step 13 represents an optimization of these parameters, for which one or more optimization techniques are preferably selected from the following group: Nelder-Mead method, Bayesian optimization, grid search, random search, gradient-free optimization, optimal experimental design, machine learning methods, greedy algorithms and evolutionary algorithms. The model of the spectrometer preferably comprises several constraints, such as a minimum bandwidth, a maximum bandwidth, a minimum wavelength and a maximum wavelength, which are taken into account when optimizing the parameters in the vector s _t (not shown).

In a step 14, the changed parameters of the model of the hardware of the spectrometer, i.e. the vector s _t (not shown), are updated for the step 07 to be carried out again for simulating the spectral measurement.

If the test in step 12 shows that the suitability parameter meets the condition, the iteratively executed steps 06 are completed. In a step 16, the optimized model of the spectrometer hardware and the chemometric model formed by the trained machine learning system are now available, so that the process as a whole is completed. To produce the

The analyzer is a spectrometer according to the optimized A model of the spectrometer hardware is to be produced and the chemometric model formed by the trained machine learning system is to be implemented in the analytical device to be manufactured. The analytical device produced in this way is ideally suited to the given analytical task and the spectrometer is far less complex than the reference spectrometer according to the gold standard, so that the analytical device can be manufactured cost-effectively.

Fig. 2 shows a flow chart of a second preferred embodiment of the method according to the invention. This second preferred embodiment, like the first preferred embodiment, serves for the technical design of an analysis device (not shown) for the spectral analysis of at least one ingredient of a sample. This second preferred embodiment also results in the analysis device to be developed being optimized for a given analysis task. The analysis device to be developed is constructed in the same way as was described for the first preferred embodiment. In contrast to the first preferred embodiment, in the second preferred embodiment the hardware of the spectrometer is physically provided as a test copy in each iteration in order to develop and optimize the spectrometer.

In a preparatory step 21, reference samples are provided which are selected according to the given analysis task. Values of concentrations of at least one ingredient of the reference sample are known for the reference samples. This is the at least one ingredient whose concentration is to be determined according to the given analysis task. These values for the reference samples Known values represent reference ingredient concentration data.

The second preferred embodiment of the method also comprises several iterative steps 22, through which a joint iterative optimization of the hardware of the spectrometer and the chemometric model in the form of the machine learning system takes place.

In an iteratively executed step 23, a test copy of the spectrometer hardware is produced. The test copy is defined by a specification s (not shown) in the form of parameters. The specification s (not shown) is preferably formed by a vector as described for the first preferred embodiment. The specification preferably also includes device parameters for signal noise and tolerances of the spectrometer, which, however, are not changed by the iteratively executed steps 22. As a result of the production, the respective test copy of the spectrometer hardware is physically available in a step 24. The first test copy of the spectrometer hardware to be produced is produced according to an initial vector So.

In a next step 26, the current test specimen of the spectrometer hardware is used to spectrally measure the reference samples, thereby obtaining spectral measurement values which form the spectral training data available in a step 27.

The spectral training data presented in step 27 are used in a step 28 to train the chemometric model formed by the machine learning system. After Each time step 28 is carried out, the machine learning system is trained and is used to predict the value of the concentration of at least one ingredient in the reference samples from the spectral training data. Deviations of the predicted values from the reference ingredient concentration data represent an error which characterizes the current test specimen of the spectrometer hardware. An objective function is therefore defined which indicates the said deviations for the respective vector s (not shown). The said deviations result in a step 29 in the value of a suitability parameter which is checked for a condition in a step 31. This condition forms the termination condition for the iteratively executed steps 22. This preferably predefined termination condition defines at what point the test specimen of the spectrometer hardware and the chemometric model formed by the machine learning system are considered sufficiently optimized to fulfill the given analysis task.

If the test in step 31 shows that the suitability parameter does not yet meet the condition, then in a step 32 modified parameters are determined for the next test example of the spectrometer hardware to be produced in order to improve the objective function described above and thus also the suitability parameter. Accordingly, in a t-th iteration 22 a vector s _t (not shown) is determined, for which preferably not only the immediately previously determined vector s _t -i (not shown) but all previously determined vectors So to s _t -i (not shown) are taken into account.

The determination of modified parameters for the next test specimen of the Hardware of the spectrometer represents an optimization of these parameters, for which one or more optimization techniques are preferably selected from the following group: Nelder-Mead method, Bayesian optimization, optimal design of experiments, machine learning methods, greedy algorithms and evolutionary algorithms. For the test specimen, several constraints are preferably defined; such as a minimum bandwidth, a maximum bandwidth, a minimum wavelength and a maximum wavelength, which are taken into account when optimizing the parameters in the vector s _t (not shown).

In a step 33, the changed parameters, i.e. the vector s _t (not shown), are updated for the re-executed step 23 of producing a test copy of the spectrometer hardware.

If the test in step 31 shows that the suitability parameter satisfies the condition, the previously iterative steps 22 are completed. In a step 34, the optimized test example of the spectrometer hardware and the chemometric model formed by the trained machine learning system are now available, so that the process as a whole is complete. To manufacture the analysis device, a spectrometer identical to the optimized test example must be manufactured and the chemometric model formed by the trained machine learning system must be implemented in the analysis device to be manufactured. The analysis device manufactured in this way is ideally suited to the given analysis task and the spectrometer is far less complex than a reference spectrometer according to the gold standard, so that the analysis device can be manufactured cost-effectively. List of reference symbols

01 Reference spectrometer

02 Reference samples

03 Recording spectra

04 Reference spectral data

05

06 iterative optimization

07 Simulating the recording

08 spectral training data

09 Training

10

11 Suitability indicator

12 Condition

13 Determining changed parameters

14 Parameter update

15

16 trained learning system and optimized model

17

18

19

20

21 reference samples

22 iterative optimization

23 Making a test copy

24 test copies

25

26 Measurement of spectra

27 spectral training data

28 Training

29 Suitability indicator 30

31 Condition

32 Determining changed parameters

33 Parameter update 34 trained learning system and optimized test specimen

Claims

Patent claims Method for the technical design of an analysis device for the spectral analysis of at least one ingredient of a sample, wherein the analysis device comprises a spectrometer for spectral measurement of the sample, wherein the analysis device is further configured to use a chemometric model formed by a trained machine learning system to determine a concentration of at least one ingredient of the sample based on spectral measurement values of the sample recorded with the spectrometer; and wherein the method comprises the following steps:

- Selection of an initial configuration of a hardware representation of the spectrometer;

- iteratively carrying out the following steps: o applying (07; 26) the respective current configuration of the representation of the hardware of the spectrometer, whereby spectral training data (08; 27) are obtained, wherein when applying (07; 26) the respective current configuration of the representation of the hardware of the spectrometer, a spectral measurement of reference samples (21) is carried out or a spectral measurement of reference samples (21) is simulated with a model of the spectrometer; o training (09; 28) the chemometric model formed by a machine learning system with the spectral training data (08; 27); o applying the chemometric model formed by the trained machine learning system to determine a value of at least one suitability parameter (11; 29) for the respective current configuration of the Representation of the hardware of the spectrometer, and o modifying (13; 32) the respective current configuration of the representation of the hardware of the spectrometer. Method according to claim 1, characterized in that after carrying out the iteratively carried out steps, that of the applied configurations of the representation of the hardware of the spectrometer is selected whose value of one of the at least one suitability parameter (11, 29) corresponds to the best possible suitability. Method according to claim 1 or 2, characterized in that in the iteratively carried out step of modifying (13; 32) the respective current configuration of the representation of the hardware of the spectrometer, the configuration of the representation of the hardware of the spectrometer is modified within a predefined framework, this iteratively carried out step being repeated until a budget predefined for this purpose is exhausted, which budget is defined by a number of iterations to be carried out and/or by a time period for carrying out the iterations. Method according to claim 1 or 2, characterized in that the steps to be carried out iteratively are repeated until the value of the suitability parameter (11, 29) for the respective current configuration of the representation of the hardware of the spectrometer has at least reached a predetermined limit value. Method according to one of claims 1 to 4, characterized in that the modification (13; 32) of the respective current configuration of the representation of the hardware of the spectrometer is carried out using at least one of the previously determined values of the at least one suitability parameter (11; 29). Method according to one of claims 1 to 5, characterized in that the representation of the hardware of the spectrometer is formed by the model of the hardware of the spectrometer, the method comprising the following further step:

- Providing reference data which comprises a series of reference spectral data (04) to which a series of reference ingredient concentration data is assigned; wherein the iteratively carried out step of applying the respective current configuration of the model of the hardware of the spectrometer comprises a simulation (07) of a spectral measurement of the sample, wherein in the simulations (07) the respective configuration of the model of the hardware of the spectrometer is applied to the reference spectral data (04), whereby spectral simulation measured values are obtained which form the spectral training data (08), so that by applying the chemometric model formed by the trained machine learning system, simulation measured values of the concentrations of the ingredient are obtained which are compared with the reference ingredient concentration data in order to determine the value of the at least one suitability parameter (11) in each case. Method according to claim 6 , characterized in that it comprises the following further steps :

- determining a manufacturing-related hardware deviation between the spectrometer hardware model and a spectrometer hardware manufactured according to the spectrometer hardware model; and

- Taking into account the manufacturing-related hardware deviation in the iteratively performed step of applying (07) the respective current configuration of the representation of the hardware of the spectrometer. Method according to claim 6 or 7, characterized in that in the iteratively performed simulation (07) of the spectral measurement of the sample, changing measurement conditions are taken into account in order to obtain extended spectral training data (08), wherein the changing measurement conditions include a distance between the sample and the spectrometer, an air humidity and/or a temperature in the area of the sample and/or the spectrometer. Method according to one of claims 1 to 5, characterized in that the representation of the hardware of the spectrometer is formed by a test copy (24) of the hardware of the spectrometer, wherein the method comprises the following further step:

- providing the reference samples (21) for which the values of the concentration of the ingredient are known, so that these values form reference ingredient concentration data; wherein the iteratively performed step of applying the respective current configuration of the test specimen (24) Measurements (26) of the reference samples with the respective current configuration of the test specimen (24) of the hardware of the spectrometer, whereby spectral measurement values are obtained which form the spectral training data (27), so that by applying the chemometric model formed by the trained machine learning system, measurement values of the concentrations of the ingredient are obtained which are compared with the reference ingredient concentration data in order to determine the value of the at least one suitability parameter (29) in each case. Method according to one of claims 1 to 9, characterized in that the respective configuration of the representation of the hardware of the spectrometer defines at least one parameter of the spectrometer; wherein the parameters specify at least one measuring channel of the spectrometer; wherein the at least one parameter is formed by a peak wavelength, a center wavelength, a mean wavelength, a bandwidth, a half-width, a curve shape parameter or a peak value of the at least one measuring channel; and wherein when modifying (13; 32) the respective current configuration of the representation of the hardware of the spectrometer, one or more of the parameters and/or a number of the measurement channels are changed. Method according to one of claims 1 to 10, characterized in that the representation of the hardware of the spectrometer defines at least one restriction which is defined by a minimum peak wavelength, by a maximum peak wavelength, by a minimum center of gravity wavelength, by a maximum Center of gravity wavelength, by a minimum mean wavelength, by a maximum mean wavelength, by a minimum bandwidth, by a maximum bandwidth, by a minimum half-width, by a maximum half-width, by a minimum curve shape parameter, by a maximum curve shape parameter, by a minimum peak value, by a maximum peak value, by a minimum wavelength difference between a wavelength parameter of one of the measuring channels and a corresponding wavelength parameter of another of the measuring channels, or by a maximum wavelength difference between a wavelength parameter of one of the measuring channels and a corresponding wavelength parameter of another of the measuring channels. Method according to one of claims 1 to 11, characterized in that the representation of the hardware of the spectrometer further defines at least one device parameter which represents signal noise, bandwidth noise, half-width noise, measurement deviation or tolerance of the spectrometer. Computer program product for the technical design of an analysis device for analyzing at least one ingredient of a sample, wherein the computer program product comprises a computer-readable storage medium having program instructions stored thereon, wherein the program instructions are executable by one or more computers or control units and cause the one or more computers or control units to carry out the method according to one of claims 1 to 12. Trained machine learning system which has been trained by a method according to one of claims 1 to 12. Analysis device for the spectral analysis of at least one ingredient of a sample, which comprises a spectrometer for spectral measurement of the sample and for applying a chemometric model formed by a trained machine learning system to determine a concentration of at least one ingredient of the sample based on spectral data recorded with the spectrometer.

Measurement values of the sample, characterized in that the spectrometer and the chemometric model formed by the trained machine learning system have emerged from the method according to one of claims 1 to 12.