Machine Learning Interatomic Potentials
Machine learning interatomic potentials (MLIPs) are methods that use machine learning models to predict interatomic interaction energies and forces, accelerating molecular dynamics simulations. They are widely applied in computational chemistry and materials science. Their core challenge is balancing predictive accuracy against physical constraints.
Timeline (last 90 days)
For the unconstrained hypothesis class, the relationship between the prediction error of MLIPs and sampling accuracy is exponentially ill-posed in terms of dimensionality: even if the prediction errors of energy and forces decay exponentially with dimensionality, the total variation distance between the learned distribution and the target distribution may still approach its maximum value.
A theoretical study proves that a model can perfectly match energies and forces on the training set yet produce completely wrong statistical behavior when sampling, that is, the distribution-level error does not decrease as local prediction accuracy improves.
This theoretical study warns against the current trend in MLIP model design of pursuing larger, more flexible, and less physically constrained models: relaxing inductive bias can improve fitting accuracy and computational efficiency but may sacrifice sampling reliability.
To ensure the sampling reliability of MLIPs, one cannot rely solely on pointwise matching of energies and forces, but needs to introduce appropriate inductive bias or constraints to guarantee the correctness of the global distribution.
All Facts (4)
For the unconstrained hypothesis class, the relationship between the prediction error of MLIPs and sampling accuracy is exponentially ill-posed in terms of dimensionality: even if the prediction errors of energy and forces decay exponentially with dimensionality, the total variation distance between the learned distribution and the target distribution may still approach its maximum value.
50%UnverifiedA theoretical study proves that a model can perfectly match energies and forces on the training set yet produce completely wrong statistical behavior when sampling, that is, the distribution-level error does not decrease as local prediction accuracy improves.
50%UnverifiedThis theoretical study warns against the current trend in MLIP model design of pursuing larger, more flexible, and less physically constrained models: relaxing inductive bias can improve fitting accuracy and computational efficiency but may sacrifice sampling reliability.
50%UnverifiedTo ensure the sampling reliability of MLIPs, one cannot rely solely on pointwise matching of energies and forces, but needs to introduce appropriate inductive bias or constraints to guarantee the correctness of the global distribution.
50%