How Is Python Transforming Materials Modeling with Machine Learning?
summary
Python workflows train atom-centered neural networks on DFT-labeled snapshots to predict energies and forces, using ASE + PyTorch for dataset generation, descriptors, autograd forces and training. A tiny proof-of-concept model on 50 distorted TiO2 structures reaches parity-plot accuracy enough to run 100-atom MD for 3 ps in one day, and claims upscaling by ~30× to approach nanosecond timescales. The approach omits electronic properties unless a learned Hamiltonian or self-consistent cycle is added; foundation models (e.g. Meta FAIRmat) are available but often require fine-tuning.
AI-written summary · about
chapters (6)
- 0:00Introduction to Python in materials
- 3:12Materials modeling methods overview
- 6:07Machine learning interatomic potentials
- 14:57Python packages in materials modeling
- 17:27Electronic properties with machine learning
- 22:32Conclusion and Q&A
related talks
11mImproving Drug Discovery with Machine Learning...
33mReproducible Molecular Simulations with the Help of Python
56mWrap up and reflection talk
41mAccelerated Discovery to Knowledge using Machine Learning
25mDeveloping Battery Materials with Python
31mThe universe as balls and springs: molecular dynamics in Python
36mAlgorithmic Alchemy: Transforming Drug Discovery with Open AI
31mInterpretability In Atomic-Scale Machine Learning