Electrocatalyst Discovery With ML Feedback and Focused Screening
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The discovery of new materials for specific applications is a lengthy and biased process, often relying on empirical methods and limited experimental data, with machine learning techniques facing challenges such as lack of universal representation and unrealistic computational models.
Innovation Solution
A machine-learning-assisted system that iteratively selects candidate samples for dark electrocatalyst and photo-electrocatalyst experiments, using a processor-based architecture to perform experiments, train models, and predict properties, allowing for the identification of compositions that satisfy predetermined performance criteria.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning techniques are used to speed up materials discovery, then productivity is improved, but reliability worsens due to lack of universal representation and unrealistic computational models
Solution Approach 1:
The system implements an iterative feedback loop where experimental results from electrochemical testing are fed back to retrain and refine the machine learning model. This continuous feedback mechanism improves prediction reliability by grounding the model in actual experimental data, while maintaining high productivity through automated iteration.
Solution Approach 2:
The system performs preliminary computational screening and filtering before experimental testing, using machine learning models to pre-identify promising candidates. This preliminary action reduces the number of materials requiring experimental validation, improving productivity while the focused experimental data subsequently refines model reliability.
2Reliability
If empirical methods are used to discover materials, then reliability is maintained through direct experimentation, but productivity deteriorates due to lengthy and tedious processes
Solution Approach 1:
The discovery process is segmented into distinct stages: computational screening, machine learning prediction, and focused experimental validation. This segmentation allows high-throughput computational methods to filter candidates while maintaining reliable experimental validation for a smaller, more manageable subset, thereby improving overall productivity without sacrificing reliability.
Solution Approach 2:
Machine learning models serve as an intermediary between computational predictions and experimental validation. This intermediary translates computational data into prioritized candidate lists for experimentation, enabling high throughput in the computational phase while maintaining reliable validation in the experimental phase, thus resolving the productivity-reliability contradiction.
3Ease of operation
If researchers focus on a few combinations of elements, then ease of operation is improved, but productivity worsens due to limited search space exploration
Solution Approach 1:
The system dynamically adjusts the search space and candidate prioritization based on accumulating experimental data. As the machine learning model is retrained with new results, it adapts to identify promising regions of the compositional space, enabling comprehensive exploration without requiring manual reconfiguration. This dynamic adaptation maintains ease of operation while dramatically improving search space coverage and productivity.
Data Source
AI summary
Methods and systems described herein concern machine-learning-assisted materials discovery. One embodiment selects a candidate sample set including a plurality of compositions and performs the following operations iteratively: (1) selects an acquisition sample set, (2) performs a dark electrocatalyst experiment or a photo-electrocatalyst experiment on the compositions in the acquisition sample set to determine one or more properties, (3) trains a machine learning model using the one or more properties, and (4) predicts, based at least in part on one or more outputs of the machine learning model, the one or more properties for one or more compositions in a test sample set including compositions on which an experiment has not yet been performed. When one or more predetermined termination criteria have been satisfied, the embodiment also identifies one or more compositions in the candidate sample set for which the one or more properties satisfy predetermined performance criteria.