Wrap up and reflection talk
summary
Argues that machine learning replaces weeks‑long DFT for small (hundreds of atoms) quantum simulations by training ML interatomic potentials on expensive force and energy labels, with two concrete routes: equivariance to SO3 and other symmetry biases, and smarter data generation. Uses energy‑predictions with predict energy then take gradient to guarantee conservation. Implements a Cartesian tensor approach as an alternative to spherical harmonics and reports it is often computationally cheaper for small l (≤4). Identifies expressivity vs optimization, tensor contraction efficiency, and runtime as open bottlenecks.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:50Molecular ML simulations
- 4:25Equivariance in chemistry
- 10:31Energy landscapes dynamics
- 16:12Reflections on equivariance
- 22:06Day one: combinatorial optimization
- 28:16Cross-community insights
- 31:35Open problems in GNNs
- 34:08Theory-practice in graph learning
- 39:49Logic and graph limits
- 44:31Neural expressivity complexity
- 45:36TCS topics for graph learning
- 50:11Algorithms in neural architectures
- 52:18Property testing streaming
- 54:30Algorithms with predictions
- 54:53Closing
related talks
31mInterpretability In Atomic-Scale Machine Learning
28mHow Is Python Transforming Materials Modeling with Machine Learning?
01h 15mImproved Machine Learning Algorithm for Predicting Ground State Properties
15mAccelerQ: Accelerating Quantum Eigensolvers With Machine Learning on Quantum Simulators
39mNetworking for Quantum: how simulations help us to design the future
33mReproducible Molecular Simulations with the Help of Python
28mNext-generation frameworks for Large-scale Machine Learning
44mQuantum machine learning: Prospects and challenges