Machine learning driven experimental design for food technology

Abstract

Techniques to generate experiment trials using artificial intelligence are disclosed. A training set for an experiment generator is continuously built up by using assessed experiment trials. The experiment generator is optimized using one of a plurality of optimization algorithms, depending on which mode experiment generator is to run in for an experiment. The mode is dependent on the experiment mode of the experiment. The experiment generator generates a batch of one or more experiment trials for the experiment. Any of the generated experiment trials may be tried or experimented by a user and may be updated with assessment data as an assessed experiment trial.

What is claimed is: 
     
         1 . A computer-implemented method comprising:
 generating, by an artificial intelligence model, a plurality of experiment trials for an experiment based on configuration data of the experiment;   wherein the artificial intelligence model includes a global optimizer configured to sample from a probability distribution of assessed experiment trials in a training dataset;   wherein the configuration data includes parameters and at least one experiment objective;   wherein each of the plurality of experiment trials includes the same set of the parameters but with varying amounts of the parameters and attempts to satisfy the at least one experiment objective;   wherein generating the plurality of experiment trials comprises:
 generating a posterior distribution of the assessed experiment trials; and 
 sampling, according to the posterior distribution, a plurality of points from within a search domain of the probability distribution of the assessed experiment trials, wherein an experiment trial of the plurality of experiment trials corresponds to a point of the plurality of points; and 
   performing at least one iteration of generating the plurality of experiment trails, wherein each of the least one iteration includes generated experiment trials refining towards the at least one experiment objective.   
     
     
         2 . The method of  claim 1 , wherein the configuration data also includes an experiment mode and at least one parameter constraint, wherein a particular representation of the parameters is defined based on the experiment mode, wherein each of the plurality of experiment trials satisfies each of the at least one parameter constraint. 
     
     
         3 . The method of  claim 1 , wherein the global optimizer is based on Bayesian Optimization. 
     
     
         4 . The method of  claim 1 , wherein the posterior distribution is generated from an optimization algorithm of a plurality of optimization algorithms, wherein the plurality of optimization algorithms includes:
 a first optimization algorithm associated with experiments that include numerical-based parameters;   a second optimization algorithm associated with experiments that include numerical-based parameters and categorical-based parameters that are not embedded in a representational space; and   a third optimization algorithm associated with experiments that include numerical-based parameters and categorical-based parameters that are embedded in the representational space.   
     
     
         5 . The method of  claim 4 , wherein the categorical-based parameters that are not embedded in the representational space are one-hot encoded and the categorical-based parameters that are embedded in the representational space are custom encoded. 
     
     
         6 . The method of  claim 4 , further comprising:
 obtaining Fourier Transform Infrared Spectroscopy (FTlR) spectra of parameters;   applying an autoencoder on the FTIR spectra to obtain encoded representations of the parameters;   wherein the categorical-based parameters that are embedded in the representational space are each associated with an encoded representation of a parameter corresponding with a respective categorical-based parameter.   
     
     
         7 . The method of  claim 1 , wherein the search domain is based the parameters, parameter constraints, and the at least one experiment objective. 
     
     
         8 . The method of  claim 1 , further comprising, prior to performing the at least one iteration, updating the training dataset with at least one experiment trial of the plurality of experiment trials and corresponding assessment data for the at least one experiment trial. 
     
     
         9 . One or more non-transitory computer-readable storage media storing one or more instructions, when executed by one or more computing devices, cause:
 generating, by an artificial intelligence model, a plurality of experiment trials for an experiment based on configuration data of the experiment;   wherein the artificial intelligence model includes a global optimizer configured to sample from a probability distribution of assessed experiment trials in a training dataset;   wherein the configuration data includes parameters and at least one experiment objective;   wherein each of the plurality of experiment trials includes the same set of the parameters but with varying amounts of the parameters and attempts to satisfy the at least one experiment objective;   wherein generating the plurality of experiment trials comprises:
 generating a posterior distribution of the assessed experiment trials; and 
 sampling, according to the posterior distribution, a plurality of points from within a search domain of the probability distribution of the assessed experiment trials, wherein an experiment trial of the plurality of experiment trials corresponds to a point of the plurality of points; and 
   performing at least one iteration of generating the plurality of experiment trails, wherein each of the least one iteration includes generated experiment trials refining towards the at least one experiment objective.   
     
     
         10 . The one or more non-transitory computer-readable storage media of  claim 9 , wherein the configuration data also includes an experiment mode and at least one parameter constraint, wherein a particular representation of the parameters is defined based on the experiment mode, wherein each of the plurality of experiment trials satisfies each of the at least one parameter constraint. 
     
     
         11 . The one or more non-transitory computer-readable storage media of  claim 9 , wherein the global optimizer is based on Bayesian Optimization. 
     
     
         12 . The one or more non-transitory computer-readable storage media of  claim 9 , wherein the posterior distribution is generated from an optimization algorithm of a plurality of optimization algorithms, wherein the plurality of optimization algorithms includes:
 a first optimization algorithm associated with experiments that include numerical-based parameters;   a second optimization algorithm associated with experiments that include numerical-based parameters and categorical-based parameters that are not embedded in a representational space; and   a third optimization algorithm associated with experiments that include numerical-based parameters and categorical-based parameters that are embedded in the representational space.   
     
     
         13 . The one or more non-transitory computer-readable storage media of  claim 12 , wherein the categorical-based parameters that are not embedded in the representational space are one-hot encoded and the categorical-based parameters that are embedded in the representational space are custom encoded. 
     
     
         14 . The one or more non-transitory computer-readable storage media of  claim 12 , wherein the one or more instructions, when executed by the one or more computing devices, further cause:
 obtaining Fourier Transform Infrared Spectroscopy (FTlR) spectra of parameters;   applying an autoencoder on the FTIR spectra to obtain encoded representations of the parameters;   wherein the categorical-based parameters that are embedded in the representational space are each associated with an encoded representation of a parameter corresponding with a respective categorical-based parameter.   
     
     
         15 . The one or more non-transitory computer-readable storage media of  claim 9 , wherein the search domain is based the parameters, parameter constraints, and the at least one experiment objective. 
     
     
         16 . The one or more non-transitory computer-readable storage media of  claim 9 , wherein the one or more instructions, when executed by the one or more computing devices, further cause, prior to performing the at least one iteration, updating the training dataset with at least one experiment trial of the plurality of experiment trials and corresponding assessment data for the at least one experiment trial. 
     
     
         17 . A computing system comprising:
 one or more computer systems comprising one or more hardware processors and storage media; and   instructions stored in the storage media and which, when executed by the computing system, cause the computing system to perform:
 generating, by an artificial intelligence model, a plurality of experiment trials for an experiment based on configuration data of the experiment; 
 wherein the artificial intelligence model includes a global optimizer configured to sample from a probability distribution of assessed experiment trials in a training dataset; 
 wherein the configuration data includes parameters and at least one experiment objective; 
 wherein each of the plurality of experiment trials includes the same set of the parameters but with varying amounts of the parameters and attempts to satisfy the at least one experiment objective; 
 wherein generating the plurality of experiment trials comprises:
 generating a posterior distribution of the assessed experiment trials; and 
 sampling, according to the posterior distribution, a plurality of points from within a search domain of the probability distribution of the assessed experiment trials, wherein an experiment trial of the plurality of experiment trials corresponds to a point of the plurality of points; and 
 
 performing at least one iteration of generating the plurality of experiment trails, wherein each of the least one iteration includes generated experiment trials refining towards the at least one experiment objective. 
   
     
     
         18 . The computing system of  claim 17 , wherein the configuration data also includes an experiment mode and at least one parameter constraint, wherein a particular representation of the parameters is defined based on the experiment mode, wherein each of the plurality of experiment trials satisfies each of the at least one parameter constraint. 
     
     
         19 . The computing system of  claim 17 , wherein the global optimizer is based on Bayesian Optimization. 
     
     
         20 . The computing system of  claim 17 , wherein the posterior distribution is generated from an optimization algorithm of a plurality of optimization algorithms, wherein the plurality of optimization algorithms includes:
 a first optimization algorithm associated with experiments that include numerical-based parameters;   a second optimization algorithm associated with experiments that include numerical-based parameters and categorical-based parameters that are not embedded in a representational space; and   a third optimization algorithm associated with experiments that include numerical-based parameters and categorical-based parameters that are embedded in the representational space.