2023.10.31 Anand Chandrasekaran, Schrödinger Inc.
Schrödinger’s AutoQSAR tool for Machine Learning can be found at: https://nanohub.org/tools/autoqsar
For details for free access on nanoHUB see: https://nanohub.org/groups/schrodinger/
Machine learning (ML) has revolutionized materials science and chemistry with the help of deep learning innovations and the availability of larger and larger datasets. Many industrial scientists want to adopt a data-driven and AI-based design approach, but they face challenges with limited datasets and complex materials that need customized feature engineering. Furthermore, typical ML methods often struggle with interpretability and generalization to new chemical domains. In this webinar, we show how Schrödinger’s tools can address these common issues by using a combination of physics-based simulation data, enterprise informatics, and chemistry-aware ML. We illustrate how this synergistic approach can transform materials innovation across a broad range of technology fields. Specifically, we will present case studies in the following areas:
• Using molecular dynamics simulations to generate features that improve the accuracy of ML models for viscosity predictions
• Building interpretable ML models to predict the ionic conductivity of Li-ion battery electrolytes
• Enhancing the performance of ML models for predicting properties such as absorption and emission wavelengths, fluorescence lifetime, and extinction coefficients of organic electronics using features derived from density functional theory
This integrated approach represents a new frontier in materials science and chemistry, combining the strengths of ML and physics-based methods.
Table of Contents:
00:00 Data-driven materials innovation: where machine learning meets physics
01:44 Machine Learning for Materials Design/Discovery at Schrödinger
03:20 Supervised Learning in Materials Science
04:33 Featurization in Diverse Materials Systems
07:09 Automated Machine Learning and Visualization in Molecular Systems
08:09 AutoQSAR for Ionic Liquids
09:22 DeepAutoQSAR: Automated Model Selection & Parameter Optimization
10:57 Case Study – Redox Flow Batteries
11:44 AutoQSAR vs DeepAutoQSAR Results
12:29 Chemical Featurization using Physics
15:16 Customized Polymer Descriptors Outperform Simple Monomers
16:39 Viscosity Dataset for Machine Learning Module
18:24 Quantitave Structure-Property Relationships (QSPR)
19:30 Impact of MD-Derived Simulation Descriptors
20:36 Impact of MD-Derived Simulation Descriptors
21:25 Machine Learning Optoelectronics Properties with DFT descriptors
21:37 Database of Optical Properties of Organic Compounds
22:35 Benchmark of DFT Descriptors
23:28 Feature Importance Analysis
24:12 Machine Learning for Volatility of Organic Molecules
24:27 Evaporation/Sublimation of Organic Molecules
25:36 Benchmarking ML Algorithms
26:13 Prediction of Pressure-Temperature Relationships
26:53 Applications of Volatility Machine Learning
27:25 Machine Learning for Inorganic 3D Crystal Structures
27:33 Transparent Conducting Oxide Band Gap ML
28:26 User Interface
28:30 DeepAutoQSAR Results
29:02 Machine Learning Property Prediction Panel
29:47 ML for Formulations
31:07 Active Learning and Genetic Optimization
31:32 Active Learning OptoElectronics Multi-Parameter Optimization (MPO)
32:02 Active Learning Workflow for OptoElectronics
33:15 Optoelectronic Genetic Optimization
34:44 Machine Learning Forcefields
34:54 Neural Network Potentials (NNPs)
35:45 Our First NN Model: Schrödinger-ANI (SANI)
36:34 QRNN: Charge-Recursive Neural Network
37:32 Bulk Properties of Liquid Electrolytes
38:32 Enterprise Informatics
38:40 Schrodinger’s Informatics Platform – LiveDesign®
39:41 Suitable for Diverse Materials and Data Types
40:12 Summary
41:03 Thank you
