Crystal Structure Prediction With Reliability-Guided Active Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The discovery of low-energy molecular crystals for drug and electronic device development is time-consuming and costly, with conventional laboratory-based methods taking over 10 years and costing over $2 billion, while computer-aided molecular crystal structure prediction (CSP) is slower and less accurate.
Innovation Solution
A machine learning model is used to predict organic crystal structures, incorporating active learning methods to improve accuracy by identifying and training on unreliable data points, generating reliable property metrics, and sorting crystal structures based on these metrics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional laboratory-based methods are used to discover molecular crystals, then the discovery process is thorough and reliable, but it takes over 10 years and costs over $2 billion
Solution Approach 1:
The patent replaces conventional laboratory-based mechanical and experimental methods with computer-aided computational methods and machine learning models. The CSP system uses algorithms to predict crystal structures, replacing wet-lab experimentation with in-silico calculations, thereby dramatically reducing both time and cost while maintaining reliability through iterative validation
Solution Approach 2:
The patent creates virtual copies of molecular crystal structures through computational modeling. Instead of physically synthesizing and analyzing every possible crystal form in the laboratory, the system generates digital representations and predictions of crystal structures, allowing researchers to screen and evaluate multiple candidates computationally before selecting a few for actual synthesis and validation
2Productivity
If computer-aided molecular crystal structure prediction (CSP) is used, then the process is faster and less expensive, but it is slower and less accurate compared to laboratory-based processes
Solution Approach 1:
The patent implements an active learning loop where the machine learning model makes predictions, uncertainty metrics identify unreliable predictions, and those cases are fed back into the training set for retraining. This iterative feedback process continuously improves model accuracy while maintaining the speed advantages of computational methods, resolving the contradiction between fast prediction and high accuracy
Solution Approach 2:
The patent employs dynamic model adaptation where the machine learning system evolves over time through active learning. The model is not static but continuously improves by identifying its own weaknesses through uncertainty quantification and selectively learning from problematic cases, allowing the system to become increasingly accurate while maintaining computational efficiency
3Productivity
If a machine learning model is used to predict all crystal structures, then the process is efficient and fast, but the accuracy decreases for unreliable data points
Solution Approach 1:
The patent applies different quality levels of prediction to different crystal structure candidates based on their individual uncertainty metrics. High-confidence predictions are generated efficiently by the machine learning model, while low-confidence cases are identified and handled separately through ground truth calculations or active learning, ensuring each prediction receives appropriate attention and accuracy
Solution Approach 2:
The patent introduces uncertainty metrics and reliability scores as intermediary indicators between the machine learning model and the final crystal structure predictions. These intermediaries allow the system to assess prediction confidence and route cases appropriately—either accepting ML predictions directly or triggering more rigorous validation processes, thereby maintaining both efficiency and reliability
Data Source
AI summary
Methods and systems for predicting crystal structures. One of the methods includes providing an indication of the one or more molecules; generating a set of crystal structures based on the indication; generating a reliability metric of a machine learning model for generating a property metric for each crystal structure in the set; calculating the property metric for each crystal structure in the set of crystal structures to generate a set of property metrics, by (i) using the machine learning model if the reliability metric for the crystal structure is within a predetermined threshold, and (ii) using a ground truth calculation of the crystal structure if the reliability metric for the crystal structure is not within the predetermined threshold; and taking an action based on the set of crystal structure indications for the one or more molecules.