In this work, MCC researchers present a machine learning method that enables molecular dynamics simulations under finite electric fields at length and time scales previously unattainable with traditional first-principles approaches. Using ARCHER2 they demonstrate its effectiveness by applying it to liquid water, achieving excellent agreement with existing experimental and computational results.
The interaction of condensed phase systems with external electric fields is of major importance in myriad processes in nature and technology. Electric fields direct the motion of cells, lead to the formation of novel materials phases on planets, induce unusual phenomena such as electrofreezing of liquids, and, of course, are essential for the storage and conversion of energy in supercapacitors, batteries and solar cells. Molecular simulation would give much needed atomistic insight into these processes but the most accurate approaches such as ab-initio or first principles-based molecular dynamics are drastically limited in scope due to their very high computational expense.
In this work we introduce a novel molecular simulation method with external electric fields denoted Perturbed Neural Network Potential Molecular Dynamics (PNNP MD). Our method boosts the accessible simulation time scale on high-performance computing platforms such as Archer2 by three to four orders of magnitude at virtually no loss in accuracy when compared to ab-initio molecular dynamics with electric fields. This advance permits calculation of dielectric properties of complex condensed phase systems that are out of reach with traditional first-principles based methods while retaining their accuracy.
We demonstrate the excellent performance of PNNP MD by calculating the dielectric constant of liquid water, a notoriously difficult problem for classical and ab-initio molecular dynamics. We obtain almost perfect agreement with experiment attesting to the accuracy of PNNP MD. We also calculate the electric-field dependent infrared spectrum of liquid water in excellent agreement with ab-initio molecular dynamics, but at a small fraction of the computational cost. Intriguingly, we find that at high electric fields liquid water becomes ice-like at room temperature, as suggested recently, and that a spectroscopic signature of this is a strong redshift in the OH stretching vibration and a strong blue-shift in the libration motion.
PNNP MD represents a major advance over previous machine learning models that have been suggested recently for the molecular dynamics simulation in the presence of electric fields. It is simple, modular, based on rigorous physical principles and systematically improvable for simulations at very high field strengths. Yet, the defining advantage over previous methods is that it does not require training against data obtained from computationally expensive ab-initio calculations in the presence of electric fields. Perhaps the most striking aspect of PNNP MD is that it predicts forces, energies and dielectric properties at ab-initio level accuracy even though the two neural network PNNP is based on were only trained on zero electric field data. Thus, PNNP MD not only interpolates but reliably extrapolates dielectric response to unseen, electrically polarized configurations.
In addition to providing a breakthrough in the simulation of atomistic systems interacting with external electric fields, PNNP MD opens the door for the simulation of systems and properties that are out of reach for ab-initio molecular dynamics. This includes the electric-field dependent ionic conductivity of liquid and solid electrolyte solutions, relevant for battery materials design, the prediction of field-dependent capacitance in e.g. supercapacitor materials and the atomistic simulation of electrochemical double layers at solid/liquid interfaces.
Outcomes:
New simulation methodology: A new machine learning method that enables molecular dynamics simulations of condensed phase systems under finite electric fields at length and time scales previously unattainable with traditional first-principles approaches.
New Knowledge: Atomistic simulation of the dielectric response of materials & prediction of dielectric properties of condensed phase systems at first-principles accuracy including dielectric constants, field-dependent vibrational spectra, field-dependent ionic mobilities and much more.
Collaborators:
Kit Joll (UCL)
Dr Philipp Schienbein (UCL, now University of Bochum, Germany)
Dr Kevin Rosso (Pacific Northwest National Laboratory, USA)
Funding:
- IMPACT PhD studentship co-sponsored by University College London and Pacific Northwest National Laboratory (PNNL) through its BES Geosciences programme (FWP 56674) supported by the U.S. Department of Energy’s Office of Science, Office of Basic Energy Sciences, Chemical Sciences, Geosciences and Biosciences Division.
- Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under project number 519139248 (Walter Benjamin Programme).
- Materials Chemistry Consortium (MCC), EPSRC (EP/L000202, EP/ R029431) for use of ARCHER2 UK National Supercomputing Service (http://www.archer2.ac.uk).
- EPSRC – UKRI ARCHER2 Pioneer Project (ARCHER2 PR17125)
Publication:
- Joll, P. Schienbein, K.M.Rosso, J. Blumberger Nat. Commun. 15, 8192 (2024) https://doi.org/10.1038/s41467-024-52491-3
Contact:
Prof Jochen Blumberger, j.blumberger@ucl.ac.uk