Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
134 tokens/sec
GPT-4o
9 tokens/sec
Gemini 2.5 Pro Pro
47 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
38 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Unified Differentiable Learning of Electric Response (2403.17207v2)

Published 25 Mar 2024 in cond-mat.mtrl-sci

Abstract: Predicting response of materials to external stimuli is a primary objective of computational materials science. However, current methods are limited to small-scale simulations due to the unfavorable scaling of computational costs. Here, we implement an equivariant machine-learning framework where response properties stem from exact differential relationships between a generalized potential function and applied external fields. Focusing on responses to electric fields, the method predicts electric enthalpy, forces, polarization, Born charges, and polarizability within a unified model enforcing the full set of exact physical constraints, symmetries and conservation laws. Through application to $\alpha$-SiO$_2$, we demonstrate that our approach can be used for predicting vibrational and dielectric properties of materials, and for conducting large-scale dynamics under arbitrary electric fields at unprecedented accuracy and scale. We apply our method to ferroelectric BaTiO$_3$ and capture the temperature-dependence and time evolution of hysteresis, revealing the underlying microscopic mechanisms of nucleation and growth that govern ferroelectric domain switching.

Definition Search Book Streamline Icon: https://streamlinehq.com
References (35)
  1. R. Resta, Macroscopic polarization in crystalline dielectrics: the geometric phase approach, Rev. Mod. Phys. 66, 899 (1994).
  2. R. Resta and D. Vanderbilt, Theory of polarization: a modern approach, in Physics of ferroelectrics: a modern perspective (Springer, 2007).
  3. R. Resta, Theory of the electric polarization in crystals, Ferroelectrics 136, 51 (1992).
  4. R. D. King-Smith and D. Vanderbilt, Theory of polarization of crystalline solids, Phys. Rev. B 47, 1651 (1993).
  5. D. Vanderbilt and R. D. King-Smith, Electric polarization as a bulk quantity and its relation to surface charge, Phys. Rev. B 48, 4442 (1993).
  6. A. K. Tagantsev, Electric polarization in crystals and its response to thermal and elastic perturbations, Phase Transit. 35, 119 (1991).
  7. D. Vanderbilt, Berry-phase theory of proper piezoelectric response, J. Phys. Chem. Solids 61, 147 (2000).
  8. N. A. Spaldin, A beginner’s guide to the modern theory of polarization, J. Solid State Chem. 195, 2 (2012).
  9. R. W. Nunes and X. Gonze, Berry-phase treatment of the homogeneous electric field perturbation in insulators, Phys. Rev. B 63, 155107 (2001).
  10. P. Umari and A. Pasquarello, Ab initio molecular dynamics in a finite homogeneous electric field, Phys. Rev. Lett. 89, 157602 (2002).
  11. J. G. Kirkwood, The dielectric polarization of polar liquids, J. Chem. Phys. 7, 911 (1939).
  12. H. Fröhlich, Theory of dielectrics: dielectric constant and dielectric loss, Clarendon Press: Oxford  (1958).
  13. R. Kubo, The fluctuation-dissipation theorem, Rep. Progr. Phys. 29, 255 (1966).
  14. M. Neumann, Dipole moment fluctuation formulas in computer simulations of polar systems, Molecular Physics 50, 841 (1983).
  15. M. Neumann and O. Steinhauser, On the calculation of the frequency-dependent dielectric constant in computer simulations, Chem. Phys. Lett. 102, 508 (1983).
  16. M. Neumann, O. Steinhauser, and G. S. Pawley, Consistent calculation of the static and frequency-dependent dielectric constant in computer simulations, Mol. Phys. 52, 97 (1984).
  17. W. Evans and J. Powles, The computer simulation of the dielectric properties of polar liquids. the dielectric constant and relaxation of liquid hydrogen chloride, Mol. Phys. 45, 695 (1982).
  18. D. McQuarrie, Statistical Mechanics; University Science books, Sausalito, CA , 222 (2000).
  19. J. Nigam, M. J. Willatt, and M. Ceriotti, Equivariant representations for molecular Hamiltonians and N-center atomic-scale properties, J. Chem. Phys. 156, 014115 (2022).
  20. M. Geiger and T. Smidt, e3nn: Euclidean neural networks,   (2022), arXiv:2207.09453 [cs.LG] .
  21. A. S. Christensen, F. A. Faber, and O. A. von Lilienfeld, Operators in quantum machine learning: Response properties in chemical space, J. Chem. Phys. 150, 064105 (2019).
  22. K. Schütt, O. Unke, and M. Gastegger, Equivariant message passing for the prediction of tensorial properties and molecular spectra, in Int. Conf. Mach. Learn., Vol. 139, edited by M. Meila and T. Zhang (PMLR, 2021) pp. 9377–9388.
  23. P. Schienbein, Spectroscopy from machine learning by accurately representing the atomic polar tensor, J. Chem. Theory Comput. 19, 705 (2023).
  24. I. Souza, J. Íñiguez, and D. Vanderbilt, First-principles approach to insulators in finite electric fields, Phys. Rev. Lett. 89, 117602 (2002).
  25. A. M. F. Neto and S. R. Salinas, The physics of ferroelectrics-a modern perspective (Springer, 2007).
  26. R. M. Pick, M. H. Cohen, and R. M. Martin, Microscopic theory of force constants in the adiabatic approximation, Phys. Rev. B 1, 910 (1970).
  27. T. Maxson and T. Szilvasi, Transferable water potentials using equivariant neural networks (2024), arXiv:2402.16204 [physics.chem-ph] .
  28. S. de Gironcoli, S. Baroni, and R. Resta, Piezoelectric properties of III-V semiconductors from first-principles linear-response theory, Phys. Rev. Lett. 62, 2853 (1989).
  29. X. Gonze, First-principles responses of solids to atomic displacements and homogeneous electric fields: Implementation of a conjugate-gradient algorithm, Phys. Rev. B 55, 10337 (1997).
  30. A. Pasquarello and R. Car, Dynamical charge tensors and infrared spectrum of amorphous SiO22{}_{2}start_FLOATSUBSCRIPT 2 end_FLOATSUBSCRIPT, Phys. Rev. Lett. 79, 1766 (1997).
  31. L. Giacomazzi, P. Umari, and A. Pasquarello, Medium-range structure of vitreous SiO22{}_{2}start_FLOATSUBSCRIPT 2 end_FLOATSUBSCRIPT obtained through first-principles investigation of vibrational spectra, Phys. Rev. B 79, 064202 (2009).
  32. S. Plimpton, Fast parallel algorithms for short-range molecular dynamics, J. Comput. Phys. 117, 1 (1995).
  33. J. P. Perdew, K. Burke, and M. Ernzerhof, Generalized gradient approximation made simple, Phys. Rev. Lett. 77, 3865 (1996).
  34. D. P. Kingma and J. Ba, Adam: A method for stochastic optimization (2017), arXiv:1412.6980 [cs.LG] .
  35. O. Marsalek and T. E. Markland, Quantum dynamics and spectroscopy of ab initio liquid water: The interplay of nuclear and electronic quantum effects, J. Phys. Chem. Lett. 8, 1545 (2017).
Citations (2)

Summary

  • The paper introduces a unified ML framework enabling simultaneous prediction of multiple electric response properties using exact differential relationships.
  • The model enforces key physical constraints with an O(3) equivariant network architecture, ensuring conservation laws and physical symmetries are preserved.
  • Applications to ferroelectric domain switching in BaTiO3 demonstrate enhanced scalability and predictive accuracy over traditional simulation methods.

Unified Differentiable Learning of Electric Response: An Overview

The paper by Falletta et al. explores the complexities of predicting the response of materials to external electric fields, a core objective in computational materials science. The authors propose a novel equivariant ML framework that seeks to unify the differential learning of generalized potentials and response properties under electric stimuli. This methodology stands out for its ability to integrate the prediction of multiple electric response properties—such as electric enthalpy, forces, polarization, Born charges, and polarizability—within a unified ML model structure, thus enforcing exact physical constraints, symmetries, and conservation laws.

Key Contributions

The core contribution of the paper is the development of a ML framework that offers significant advantages over traditional methods by learning from a generalized potential energy framework. Key contributions include:

  • Unified Model for Electric Response: The framework enables the simultaneous prediction of multiple response properties by using exact differential relationships between the generalized potential energy and response functions. This is not typically feasible with conventional models where dielectric properties and force fields are often learned separately.
  • Enforcement of Physical Constraints: The model ensures compliance with momentum conservation, the acoustic sum rule for Born charges, and the conservation of electric enthalpy. This is achieved by leveraging equivariant neural networks that respect O(3) symmetry, ensuring that fundamental physical laws are not violated.
  • Application to Ferroelectric Dynamics: The framework is applied to paper ferroelectric domain switching using BaTiO3_3. The authors capture the temperature-dependence and time evolution of hysteresis, detailing the microscopic mechanisms of nucleation and domain growth that characterize the ferroelectric behavior.

Methodology

The authors employ an equivariant neural network model that treats both the atomic positions and external fields as inputs. Key features of the methodology include:

  • Differential Learning Approach: By differentiating the generalized potential energy, the model calculates response properties directly, offering a robust way to learn complex, coupled dependencies on multiple parameters.
  • Equivariant Network Architecture: The use of equivariant neural networks enables data efficiency and scalability, critical for handling large atomistic simulations efficiently.
  • Sobolev Training Framework: The training leverages the Sobolev framework, minimizing a loss function composed of differences between predicted observables and their corresponding gradients, thus ensuring high fidelity in learning response properties.

Implications and Future Directions

Practically, this unified approach offers a scalable method able to surpass the time and length scale limitations usually encountered with quantum mechanical simulations. By modeling electric responses at an atomistic level, it facilitates the exploration of complex dielectric and ferroelectric phenomena in large-scale systems. This has potential applications in the design and optimization of materials for electronic devices, where understanding and control of electric response is essential.

Theoretically, the framework underscores the potential of machine learning to integrate with and augment traditional physics-based computational methods, suggesting future developments in generalized potential modeling and invoking higher-order response interactions.

Conclusion

Falletta et al.'s paper provides significant insights into the efficient, comprehensive modeling of electric response through a singular, unified ML architecture. It not only achieves impressive numerical agreement with existing first-principles methods but also extends the capabilities of such methods to more extensive and practical systems. This work sets a promising precedent for future exploration in the domain of machine learning-aided materials science, where the nuanced interplay between materials' structure and response can be more effectively harnessed for technological innovation.