Learning neural network potentials from experimental data via Differentiable Trajectory Reweighting.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Neural network (NN) potentials are increasingly used in molecular dynamics (MD) simulations.
• Current NN potentials are primarily trained using bottom-up approaches with quantum mechanical data.
• Top-down approaches learning from experimental data face significant computational challenges.

Purpose of the Study:

• To develop a novel method for training neural network potentials using experimental data.
• To overcome the numerical and computational hurdles in top-down NN potential learning.
• To enable the use of diverse experimental observables for NN potential training.

Main Methods:

• The Differentiable Trajectory Reweighting (DiffTRe) method is presented.
• DiffTRe bypasses direct differentiation through MD simulations for time-independent observables.
• Thermodynamic perturbation theory is leveraged to stabilize gradient computation and accelerate learning.

Main Results:

• DiffTRe achieves approximately two orders of magnitude speed-up in gradient computation.
• Effective NN potentials were learned for atomistic diamond and coarse-grained water models.
• The method successfully incorporated thermodynamic, structural, and mechanical experimental data.
• DiffTRe generalizes existing coarse-graining methods, such as iterative Boltzmann inversion.

Conclusions:

• DiffTRe offers an efficient and effective approach for top-down learning of neural network potentials.
• The method facilitates the integration of experimental data, crucial when high-quality quantum mechanical data is unavailable.
• DiffTRe represents a significant advancement in developing more accurate and versatile NN potentials.