Delta Machine Learning for Predicting Dielectric Properties and Raman Spectra

Abstract: Raman spectroscopy is an important characterization tool with diverse applications in many areas of research. We propose a machine learning method for predicting polarizabilities with the goal of providing Raman spectra from molecular dynamics trajectories at reduced computational cost. A linear-response model is used as a first step and symmetry-adapted machine learning is employed for the higher-order contributions as a second step. We investigate the performance of the approach for several systems including molecules and extended solids. The method can reduce training set sizes required for accurate dielectric properties and Raman spectra in comparison to a single-step machine learning approach.

• The paper proposes a Δ-ML approach that utilizes linear response models followed by machine learning refinement, achieving near-unity R² for polarizability predictions.
• The methodology leverages kernel-based techniques and symmetry-adapted descriptors like SOAP to significantly reduce computational costs relative to full DFPT calculations.
• Results demonstrate robust performance across systems such as SiO₂, AlN, and NaCl, enabling accurate and efficient predictions of Raman spectra and dielectric responses.

Delta Machine Learning for Predicting Dielectric Properties and Raman Spectra

Introduction

This paper proposes a novel delta machine learning ($\Delta$-ML) approach designed to enhance the predictive capabilities of Raman spectra through the efficient estimation of dielectric properties. Raman spectra serve as a pivotal tool in analyzing the dynamical behavior of molecular and solid-state systems at finite temperatures. However, these computations are traditionally resource-intensive, constrained by the necessity of extensive molecular dynamics (MD) trajectory sampling. The presented $\Delta$-ML method circumvents these computational burdens by utilizing a two-step process, combining classical linear-response models (LRMs) with symmetry-adapted machine learning techniques, thereby streamlining data requirements and improving prediction accuracy.

Methodology

The $\Delta$-ML method enhances the prediction of polarizabilities, the cornerstone of Raman spectra calculation, by first applying a computationally economical LRM. This model acts as a precursor, capturing baseline dielectric responses, which are successively refined through machine learning algorithms adapted for tensorial properties.

1. Linear Response Model (LRM):
   • The LRM is conceptualized via the Taylor expansion of a polarizability tensor component relative to atomic displacements (Equation 1). This expansion isolates the constant term $\alpha(x_0)$ through initial density functional perturbation theory (DFPT) calculations and further involves the first-order derivatives, which are extracted through additional DFPT analyses on displaced atomic coordinates.

• Symmetrical considerations reduce the computational overhead by limiting the number of necessary DFPT calculations. These calculations yield a fundamental polarizability profile against which machine learning techniques can be applied.

  Figure 1: LRM predictions of the $\alpha_{xx}$ component of the polarizability tensor in \ch{SiO2}.

1. Machine Learning Enhancements:
   • The ML component leverages kernel-based methods, employing descriptors that encapsulate atomic configuration symmetries facilitated through smooth overlap of atomic positions (SOAP) and à la carte enhancements like $\lambda$-SOAP for covariant considerations.

• The ML algorithm is trained on DFPT-derived polarizability components by stripping LRM predictions, thereby focusing the learning algorithm on residual errors rather than absolute values.

  Figure 2: Scatterplot comparing direct-ML predictions of polarizability components in \ch{SiO2}.

Results and Performance

The $\Delta$-ML model demonstrates robust predictive performance across various molecular and solid-state configurations, showing significant improvements over direct machine learning models notably in reducing the necessary training set sizes while maintaining prediction accuracy.

• Comparison Metrics:

Validation against DFPT reference datasets revealed that the $\Delta$-ML model regularly achieves a coefficient of determination ($R^2$) close to unity with considerably smaller training sets, improving computational efficiency.

Figure 3: Performance metrics for direct ML predictions in \ch{SiO2}.

• Diverse System Analysis:

Systems including \ch{SiO2}, AlN, and NaCl illustrated the model's versatility across varying interatomic interactions and complexities due to LO/TO splitting effects. Notably, the LRM-model synergy exhibited resilience even in challenging non-first-order Raman inactive materials like NaCl.

Implications and Future Work

The proposed method offers a substantial leap forward in the computational prediction of dielectric properties, with implications extending well beyond Raman spectra, potentially including infrared spectra and transport coefficient calculations. The $\Delta$-ML approach provides a framework ready for adaptation with advanced physical models, alternative descriptors, and refined hyperparameter tuning strategies.

Figure 4: Raman spectrum for \ch{SiO2} computed from $\Delta$-ML predictions.

Conclusion

In summary, $\Delta$-ML stands as a promising technique enhancing the accuracy and efficiency of Raman spectra predictions by bridging conventional response models with sophisticated machine learning for real-time molecular simulations. It offers high predictive accuracy with reduced data, resource requirements, and computational expenses. Future explorations may determine broader applications in fields requiring real-time monitoring and predictions of dynamic atomic interactions under varying conditions.