Northwestern Chemistry welcomes Bingqing Cheng from the University of California, Berkeley, hosted by Songi Han.
Machine learning the physics of fluctuating matter
Quantum mechanics supplies the microscopic laws of chemistry, yet much of what we measure, such as solubility, phase equilibria, interfacial structure, and response to electric fields, emerges from the collective, thermally fluctuating motion of many atoms, far beyond the reach of direct electronic-structure calculation.
In this talk I will argue that machine learning, combined with statistical mechanics, can do more than accelerate simulation: it can learn the transferable physical objects from which these observables follow. I will present three examples from my group. Latent Ewald Summation adds long-range electrostatics to machine-learning interatomic potentials using only energies and forces, yielding field-responsive models that predict infrared and Raman spectra, ferroelectric hysteresis, and the structure and capacitance of electrified metal/water interfaces. The S0 method extracts chemical potentials from the structure factors of equilibrium simulations, enabling predictions of solubilities, phase separation in planetary interiors, and ion activities in complex electrolytes. A classical density functional theory whose free-energy functional is learned from equilibrium density fields recovers equations of state, phase coexistence, solvent- mediated forces, and adsorption in porous media, none of which were fitted. Together, these tools extend first-principles prediction to the quantities chemists measure, in systems where empirical parameterization has usually been necessary.
Jillian Wilson
Email