Machine Learning Prediction of the Experimental Transition Temperature of Fe(II) Spin-Crossover Complexes

Vyshnavi Vennelakanti, Irem B. Kilic, Gianmarco Terrones, Chenru Duan, Heather J. KulikThe Journal of Physical Chemistry A 2023 · DOI

Abstract

Spin-crossover (SCO) complexes are materials that exhibit changes in the spin state in response to external stimuli, with potential applications in molecular electronics. It is challenging to know a priori how to design ligands to achieve the delicate balance of entropic and enthalpic contributions needed to tailor a transition temperature close to room temperature. We leverage the SCO complexes from the previously curated SCO-95 data set to train three machine learning (ML) models for transition temperature ($T_{1/2}$) prediction using graph-based revised autocorrelations as features. We perform feature selection using random forest-ranked recursive feature addition (RF-RFA) to identify the features essential to model transferability. Of the ML models considered, the full feature set RF and recursive feature addition RF models perform best, achieving moderate correlation to experimental $T_{1/2}$ values. We then compare ML $T_{1/2}$ predictions to those from three previously identified best-performing density functional approximations (DFAs), finding that the ML models predict $T_{1/2}$ more accurately than the best-performing DFAs. In addition, we study ML model predictions for a set of 18 SCO complexes for which only estimated $T_{1/2}$ values are available. Upon excluding outliers from this set, the RF-RFA RF model shows a strong correlation to estimated $T_{1/2}$ values with a Pearson's $r$ of 0.82. Overall, our study demonstrates slightly superior performance of ML models in comparison with some of the best-performing DFAs.

Introduction

Spin-crossover (SCO) complexes are the "chameleons" of the molecular world. These mononuclear octahedral Fe(II) complexes can switch between low-spin (LS) and high-spin (HS) states when triggered by temperature, pressure, or light. This bistability makes them prime candidates for next-generation molecular memory, sensors, and spintronics.

However, designing these materials presents a "Goldilocks" problem. For practical applications, the transition temperature ($T_{1/2}$) must be precisely tuned, typically near room temperature. This requires a delicate balance between enthalpy and entropy ($\\Delta H / \\Delta S$). While Density Functional Theory (DFT) is the standard tool for molecular modeling, it struggles significantly with $T_{1/2}$ prediction. DFT can often identify if a molecule might switch, but it is notoriously inaccurate at predicting the specific temperature, often producing errors of hundreds of Kelvin due to the compounding effects of small errors in vibrational entropy calculations.

Methodology

To overcome the limitations of physics-based simulations, this study shifts toward a data-driven machine learning framework.

Graph-Based Descriptors

Instead of relying on 3D coordinates—which are computationally expensive to obtain for novel molecules—the models utilize Revised Autocorrelations (RACs). These are graph-based descriptors that represent the molecule as a social network of atoms. RACs encode heuristic properties such as electronegativity, nuclear charge, and covalent radius across various bond distances (depths), allowing the model to capture both local and global molecular environments.

Model Architecture and Feature Selection

Three primary modeling approaches were evaluated:

  1. Random Forest (RF): Trained on the full feature set of 150 RACs.
  2. RF with Recursive Feature Addition (RF-RFA): A pruned model that identifies the most essential features to ensure transferability and interpretability.
  3. Kernel Ridge Regression (KRR): Utilizing the features selected via RF-RFA.

Datasets

The models were trained and validated using the SCO-95 dataset, which includes 95 mononuclear octahedral Fe(II) complexes. The data was partitioned into a training/test set of 76 complexes with known experimental $T_{1/2}$ values and a "blind" set of 18 complexes where $T_{1/2}$ was estimated from crystallization temperatures. Results were benchmarked against 28 different DFT functionals, specifically the top performers: modB3LYP, TPSSh, and M06-L.

Key Insights

The Importance of Global Features

A significant finding of this research is the dominance of "metal-distal" features. The most critical predictors for $T_{1/2}$ were not the atoms in the immediate first coordination sphere of the Iron center. Instead, global features—atoms further away in the ligand—proved more influential. This suggests that the outer shell of the molecule plays a massive role in tuning the transition via long-range electronic effects.

Electronic vs. Geometric Factors

Electronic properties, including nuclear charge and electronegativity, accounted for over 55% of the models' decision-making weight. These factors were found to be more influential than simple geometric factors like connectivity or atom size, highlighting the electronic nature of the spin-state transition.

Results and Discussion

The ML models demonstrated a clear advantage over traditional quantum chemistry methods:

  • Predictive Accuracy: The best ML models (RF and RF-RFA) achieved a Mean Absolute Error (MAE) of approximately 40 K.
  • ML vs. DFT: In contrast, the best-performing DFT functional (M06-L) yielded an MAE of 162 K, while other functionals exceeded 500 K.
  • Blind Set Validation: On a refined set of 9 blind complexes (excluding outliers and complex two-step transitions), the RF-RFA model showed a strong correlation with experimental estimates (Pearson’s $r = 0.82$). Under the same conditions, DFT showed virtually zero correlation, failing to capture the trend of the transition temperatures.

Conclusion

This study demonstrates that machine learning, specifically using graph-based descriptors and Random Forest architectures, provides a more accurate and efficient route to predicting SCO transition temperatures than traditional DFT. While the current models are limited by a relatively small training set and a focus on single-step transitions, they represent a significant step forward in the rational design of molecular switches. As larger, curated datasets of SCO complexes become available, the predictive power of these ML frameworks is expected to improve, further enabling the engineering of materials for molecular electronics.

Summary updated

Abstract

Spin-crossover (SCO) complexes are materials that exhibit changes in the spin state in response to external stimuli, with potential applications in molecular electronics. It is challenging to know a priori how to design ligands to achieve the delicate balance of entropic and enthalpic contributions needed to tailor a transition temperature close to room temperature. We leverage the SCO complexes from the previously curated SCO-95 data set [Vennelakanti et al. <i>J. Chem. Phys</i>. <i>159,</i> 024120 (<b>2023</b>)] to train three machine learning (ML) models for transition temperature (<i>T</i><sub>1/2</sub>) prediction using graph-based revised autocorrelations as features. We perform feature selection using random forest-ranked recursive feature addition (RF-RFA) to identify the features essential to model transferability. Of the ML models considered, the full feature set RF and recursive feature addition RF models perform best, achieving moderate correlation to experimental <i>T</i><sub>1/2</sub> values. We then compare ML <i>T</i><sub>1/2</sub> predictions to those from three previously identified best-performing density functional approximations (DFAs) which accurately predict SCO behavior across SCO-95, finding that the ML models predict <i>T</i><sub>1/2</sub> more accurately than the best-performing DFAs. In addition, we study ML model predictions for a set of 18 SCO complexes for which only estimated <i>T</i><sub>1/2</sub> values are available. Upon excluding outliers from this set, the RF-RFA RF model shows a strong correlation to estimated <i>T</i><sub>1/2</sub> values with a Pearson's <i>r</i> of 0.82. In contrast, DFA-predicted <i>T</i><sub>1/2</sub> values have large errors and show no correlation to estimated <i>T</i><sub>1/2</sub> values over the same set of complexes. Overall, our study demonstrates slightly superior performance of ML models in comparison with some of the best-performing DFAs, and we expect ML models to improve further as larger data sets of SCO complexes are curated and become available for model training.
BibTeX
@article{vennelakanti2023machine,
  author = {Vyshnavi Vennelakanti and Irem B. Kilic and Gianmarco Terrones and Chenru Duan and Heather J. Kulik},
  title = {{Machine Learning Prediction of the Experimental Transition Temperature of Fe(II) Spin-Crossover Complexes}},
  journal = {The Journal of Physical Chemistry A},
  year = {2023},
  doi = {10.1021/acs.jpca.3c07104},
}
Figure