# US 2024/0203537 A1 · Hybrid machine learning methods of training and using models to predict formulation properties

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2024/0203537 A1
- Document ID: US20240203537A1
- Status: Published application
- Kind: A1
- Application number: 18/085,318
- Filed: 2022-12-20
- Published: 2024-06-20
- Earliest claimed: 2022-12-20
- Assignees: Dow Global Technologies LLC
- Applicants: Dow Global Technologies LLC
- Inventors: Fabio Aguirre Vargas; Sukrit Mukhopadhyay; Jerome Claracq; Bart Rijksen; Valeriy V. Ginzburg; Huikuan Chao; Paul Cookson; Alix Schmidt; Shachit Shankaran Lyer
- Web page: https://patentgrep.com/patent/US20240203537A1

## CPC

- G16C 20/30: Prediction of properties of chemical compounds, compositions or mixtures
- G06N 5/01: Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
- G16C 20/70: Machine learning, data mining or chemometrics
- G16C 60/00: Computational materials science, i.e. ICT specially adapted for investigating the physical or chemical properties of materials or phenomena associated with their design, synthesis, processing, characterisation or utilisation
- G06N 20/20: Ensemble learning
- G16C: COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

Methods include training a machine learning module to predict one or more target product properties for a prospective chemical formulation, including (a) constructing or updating a training data set from one or more variable parameters; (b) performing feature selection on the training data set; (c) building one or more machine learning models using one or more model architectures; (d) validating the one or more machine learning models; (e) selecting at least one of the one or more machine learning models and generating prediction intervals; (g) interpreting the one or more machine learning models; and (h) determining if the one or more target product properties calculated are acceptable and deploying one or more trained machine learning models, or optimizing the one or more machine learning models by repeating steps (b) to (g). Methods also include application of trained machine learning modules to predict formulation properties from prospective data.

## Claims

### Claim 1

A method of training a machine learning module to predict one or more target product composition properties for a prospective polyurethane chemical formulation, comprising:
(a) constructing or updating a training data set from variable parameters comprising at least one selected from each of:
(i) one or more of chemical components and process conditions,
(ii) one or more polyurethane specific descriptors, comprising expressions describing a correlation within a chemical system that provide information regarding behavior of the chemical system, and
(iii) the one or more target product composition properties;
(b) performing feature selection on the training data set to determine a subset of driving variable parameters for calculating the one or more target product properties, wherein the subset of driving variable parameters comprises at least one polyurethane specific descriptor;
(c) building one or more machine learning models using one or more model architectures and analyze the training data set and subset of driving variable parameters used in calculating the one or more target product properties;
(d) validating the one or more machine learning models by inputting a testing data set and determining an associated error for the one or more machine learning models calculating the one or more target product properties;
(e) selecting at least one of the one or more machine learning models and generating prediction intervals for the one or more machine learning models;
(f) interpreting the one or more machine learning models and analyzing the one or more target product properties calculated by the one or more machine learning models;
(g) determining if the one or more target product properties calculated by the one or more machine learning models are acceptable based on pre-defined criteria; and
(h) deploying one or more trained machine learning models;
wherein the method is performed by one or more processors.

### Claim 2 (depends on 1)

The method of claim 1, wherein the prediction intervals are generated using a bootstrapping or ensemble method.

### Claim 3 (depends on 1)

The method of claim 1, wherein the associated error of (d) is determined by one or more of MAPE, R2, and RMSE.

### Claim 4 (depends on 1)

The method of claim 1, wherein the one or more model architectures is selected from one or more of selected from neural networks, symbolic regression, decision trees, random forests, boosted trees, linear regression, partial least squares regression, support vector machines, multilayer perceptron, Bayesian networks, and hidden Markov models.

### Claim 5 (depends on 1)

The method of claim 1, wherein the model outputs are evaluated using Shapley additive explanation (SHAP), local interpretable model-agnostic explanation (LIME), or Variable Importance to the Projection (VIP).

### Claim 6 (depends on 1)

The method of claim 1, wherein (f) further comprises pruning the one or more machine learning models by removing variable parameters that do not contribute to the one or more target product properties calculated.

### Claim 7 (depends on 1)

The method of claim 1, wherein (f) further comprises introducing additional variable parameters into the one or more machine learning models and repeating steps (b) to (f).

### Claim 8 (depends on 1)

The method of claim 1, wherein the one or more trained machine learning models have a prediction accuracy as calculated by R2 of at least 0.75.

### Claim 9

A method of determining one or more target product composition properties from a prospective polyurethane chemical formulation, comprising:
inputting the prospective chemical formulation into a trained machine learning module, wherein training the trained machine learning module comprises:
(a) constructing or updating a training data set from variable parameters comprising at least one selected from each of:
(i) one or more of chemical components and process conditions,
(ii) one or more polyurethane specific descriptors, comprising expressions describing a correlation within a chemical system that provide information regarding behavior of the chemical system, and
(iii) the one or more target product composition properties;
(b) performing feature selection on the training data set to determine a subset of driving variable parameters for calculating the one or more target product properties, wherein the subset of driving variable parameters comprises at least one polyurethane specific descriptor;
(c) building one or more machine learning models using one or more model architectures and analyze the training data set and subset of driving variable parameters used in calculating the one or more target product properties;
(d) validating the one or more machine learning models by inputting a testing data set and determining an associated error for the one or more machine learning models calculating the one or more target product properties;
(e) selecting at least one of the one or more machine learning models and generating prediction intervals for the one or more machine learning models using a bootstrapping method;
(f) interpreting the one or more machine learning models and analyzing the one or more target product properties calculated by the one or more machine learning models;
(g) determining if the one or more target product properties calculated by the one or more machine learning models are acceptable and deploying one or more trained machine learning models;
inputting a selection of one or more target product properties to be predicted into the machine learning module; and
outputting the one or more predicted target product properties;
wherein the method is performed by one or more processors.

### Claim 10 (depends on 9)

The method of claim 9, further comprising accepting or rejecting a product produced from the prospective chemical formulation and/or modifying a process generating the product based on the one or more predicted target product properties.

### Claim 11 (depends on 9)

The method of claim 9, wherein the one or more trained machine learning modules have a prediction accuracy as calculated by R2 of at least 0.75.

### Claim 12 (depends on 1)

The method of claim 1, further comprising optimizing the one or more machine learning models by repeating steps (b) to (f).

### Claim 13 (depends on 9)

The method of claim 9, further comprising optimizing the one or more machine learning models by repeating steps (b) to (f).

### Claim 14 (depends on 9)

The method of claim 9, wherein performing feature selection on the training data set to determine a subset of driving variable parameters comprises selecting at least one chemical classification from the one or more chemical components.

### Claim 15 (depends on 1)

The method of claim 1, wherein the polyurethane specific descriptors comprise one or more of a group of polyurethane specific descriptors including NCO index, OH number, functionality, equivalent weight, molecular weight, molar volume, computed polar surface area, number of hydrogen bond acceptors, partition coefficient, water content, and solids content.

### Claim 16 (depends on 9)

The method of claim 9, wherein the polyurethane specific descriptors comprise one or more of a group of polyurethane specific descriptors including NCO index, OH number, functionality, equivalent weight, molecular weight, molar volume, computed polar surface area, number of hydrogen bond acceptors, partition coefficient, water content, and solids content.
