Predicting phosphorescence energies and inferring wavefunction localization with machine learning

Andrew E. Sifain, Levi Lystrom, Richard A. Messerly, Justin S. Smith, Benjamin Nebgen, Kipton Barros, Sergei Tretiak, Nicholas Lubbers, Brendan J. Gifford

Introduction

The development of efficient phosphorescent materials is a cornerstone of modern optoelectronics, with critical applications in organic light-emitting diodes (OLEDs) and solar energy harvesting. A key parameter in designing these materials is the singlet-triplet energy gap ($\Delta E$), which determines the emission color and the efficiency of the phosphorescence process. Traditionally, researchers have relied on ab initio quantum chemistry methods, such as Density Functional Theory (DFT), to calculate these gaps. While accurate, DFT is computationally intensive, often requiring hours or days for a single complex molecule. This bottleneck severely limits the high-throughput screening of potential candidates.

Machine learning (ML) has emerged as a promising alternative, offering the potential to predict molecular properties at a fraction of the computational cost. However, standard ML architectures for chemistry often struggle with electronic excitations. This paper introduces a novel architecture, HIP-loc, which incorporates "localization layers" to accurately model the spatially localized nature of phosphorescence, providing a significant leap in both predictive accuracy and physical interpretability.

The Extensivity Problem in Molecular ML

Most state-of-the-art ML models for molecular systems are designed to predict ground-state properties, such as total potential energy. These models typically rely on the principle of extensivity: the assumption that the total energy of a system is the sum of the energies of its constituent parts. In this framework, if a molecule's size is doubled, its predicted energy roughly doubles.

While appropriate for ground-state thermodynamics, extensivity fails for electronic excitations. Phosphorescence is an inherently localized phenomenon; the transition between singlet and triplet states often occurs within a specific functional group or a localized aromatic region, while the rest of the molecule remains a "spectator." Standard atom-centered neural networks (ACNNs) tend to distribute the predicted energy gap across all atoms in a molecule. When these models are trained on small molecules and then applied to larger ones, this summation approach leads to massive overestimations of the energy gap, as the model incorrectly adds up contributions from spectator atoms that do not participate in the transition.

Methodology: The HIP-loc Architecture

To address the failure of extensivity, the researchers developed HIP-loc, an evolution of the Hierarchical Interacting Particle Neural Network (HIPNN). The core innovation is the introduction of localization layers.

Instead of a simple global sum of atomic contributions, HIP-loc employs a weighting mechanism inspired by the attention layers used in Large Language Models (LLMs). The model learns to assign an "excitation propensity" to each atom. These propensities are passed through a softmax function to generate a set of weights ($w_i$) that sum to one:

$$\Delta E = \sum_{i} w_i \epsilon_i$$

where $\epsilon_i$ represents the local energy contribution from atom $i$. By using a weighted average rather than a sum, the model can effectively "zero out" the contributions of atoms far from the site of the electronic transition. This allows the network to isolate the chemical environment most determinative of the energy gap, regardless of the overall size of the molecule.

Data and Training

The model was trained on a massive dataset of approximately 702,000 organic structures derived from the GDB13 database, consisting of small molecules with up to 12 heavy atoms. To test the model's ability to generalize to real-world materials, the researchers curated an "extensibility set" of 35 unique large phosphorescent compounds (comprising 912 conformers) known from experimental literature. These molecules, containing up to 31 heavy atoms, are significantly larger and more complex than anything in the training set, representing a rigorous test of the model's physical foundations.

Results and Physical Insights

The results demonstrate that localization is the "secret sauce" for predicting excited-state properties. On the large-molecule extensibility set, the standard HIPNN model (which assumes extensivity) failed dramatically, yielding a Root-Mean-Square Error (RMSE) of 53 kcal/mol. In contrast, HIP-loc reduced this error to 13 kcal/mol—a four-fold improvement.

Beyond numerical accuracy, HIP-loc exhibited superior physical fidelity. Standard models often produced "unphysical" predictions, such as negative energy gaps (implying a triplet state lower than the ground state), which are impossible for the molecules studied. HIP-loc consistently maintained positive, physically plausible gaps.

Perhaps the most remarkable finding was the emergent physics within the model. Although HIP-loc was trained only on total energy labels and received no explicit electronic information, the learned atomic weights ($w_i$) showed a strong correlation with the ab initio spin density of the singlet-triplet transition. The model independently "discovered" the physical location of the triplet excitation, identifying the specific aromatic rings or functional groups where the electronic transition was concentrated.

Conclusion and Future Outlook

HIP-loc represents a paradigm shift in how we apply machine learning to excited-state chemistry. By moving away from the assumption of extensivity and embracing localization, the model achieves a million-fold speedup over DFT while maintaining the accuracy needed for material discovery.

The implications of this work extend beyond phosphorescence. The localization layer approach is a general solution for modeling any non-extensive, localized phenomenon. Future applications could include predicting the properties of charged radicals, modeling charge transport in organic semiconductors, or identifying active sites in complex catalytic systems. As we move toward "end-to-end" ML tools that can also perform geometry optimizations, architectures like HIP-loc will be essential for the rapid, accurate design of next-generation optoelectronic materials.

Summary updated

Abstract

Phosphorescence is commonly utilized for applications including light-emitting diodes and photovoltaics. Machine learning (ML) approaches trained on <i>ab initio</i> datasets of singlet-triplet energy gaps may expedite the discovery of phosphorescent compounds with the desired emission energies. However, we show that standard ML approaches for modeling potential energy surfaces inaccurately predict singlet-triplet energy gaps due to the failure to account for spatial localities of spin transitions. To solve this, we introduce localization layers in a neural network model that weight atomic contributions to the energy gap, thereby allowing the model to isolate the most determinative chemical environments. Trained on the singlet-triplet energy gaps of organic molecules, we apply our method to an out-of-sample test set of large phosphorescent compounds and demonstrate the substantial improvement that localization layers have on predicting their phosphorescence energies. Remarkably, the inferred localization weights have a strong relationship with the <i>ab initio</i> spin density of the singlet-triplet transition, and thus infer localities of the molecule that determine the spin transition, despite the fact that no direct electronic information was provided during training. The use of localization layers is expected to improve the modeling of many localized, non-extensive phenomena and could be implemented in any atom-centered neural network model.
BibTeX
@inproceedings{sifain2021predicting,
  author = {Andrew E. Sifain and Levi Lystrom and Richard A. Messerly and Justin S. Smith and Benjamin Nebgen and Kipton Barros and Sergei Tretiak and Nicholas Lubbers and Brendan J. Gifford},
  title = {{Predicting phosphorescence energies and inferring wavefunction localization with machine learning}},
  booktitle = {Chemical},
  year = {2021},
  doi = {10.1039/d1sc02136b},
}
Figure