Predicting Spectroscopic Properties of Solvated Nile Red with Automated Workflows for Machine Learned Interatomic Potentials

TL;DR

The paper tackles predicting solvated spectroscopy with high accuracy while curbing computational cost by using ESTEEM, an automated workflow that generates and refines MLIPs via active learning to reproduce TD-DFT electronic excitations. It systematically evaluates how solute–solvent cluster size, MLIP architecture (SH vs MH), and energy-gap versus ground/excited-state subtraction strategies affect UV–Vis absorption and emission spectra for solvatochromic Nile Red in multiple solvents. The results show that incorporating larger solvent environments ($R_ ext{carve} \approx 5.0~\mathrm{\AA}$) and using delta-energy or energy-gap models (especially SH-EG) yields spectra with accuracy approaching ground-truth TD-DFT while maintaining computational feasibility, with spectral convergence observed at moderate cluster sizes. Collectively, ESTEEM provides a scalable, transferable framework for environment-aware spectral predictions, enabling routine theoretical spectroscopy that accounts for vibronic effects and solvent structure.

Abstract

Machine Learned Interatomic Potentials (MLIPs) offer a powerful combination of abilities for accelerating theoretical spectroscopy calculations utilising both ensemble sampling and trajectory post-processing for inclusion of vibronic effects, which can be very challenging for traditional ab initio MD approaches. We demonstrate a workflow that enables efficient generation of MLIPs for the solvatochromic dye nile red system, in a variety of solvents. We use iterative active learning techniques to make this process as efficient as possible in terms of number and size of Density Functional Theory (DFT) calculations. Additionally, we compare the efficacy of various methodologies: generating distinct MLIPs for each adiabatic state, using one ground state MLIP in combination with delta-ML of excitation energies, and using a three-headed multiheaded ML model. To evaluate the validity of the resulting models, we compare predicted absorption and emission spectra to experimental spectra. We found that the incorporation of larger solvent systems into training data, and the use of delta models to predict the excitation energies, enables the accurate and affordable prediction of UV-Vis spectra with accuracy equivalent to the ground truth method, which is time-dependent DFT in this case.

Figures (14)

• Figure 1: Visualisation of workflow in ESTEEM. Rounded rectangles represent ESTEEM's 'tasks', cylinders represent datasets produced, ovals are trained ML models, and illustrative pages represent input choices and output data. The iterative loop within the Active Learning section is repeated 2-3 times, with selection of a set of new geometries for DFT calculations based on evaluating the standard deviation of energy predictions for a cluster over a committee of ML models.
• Figure 2: Sample standard deviations in spectra peak maxima predicted using each of the 5 SH-EG models in the committees in model set 2 to infer energy gaps for NR in ethanol clusters carved with radii of 0.0, 2.5, 5.0 and 7.5 Å.
• Figure 3: Absorption (blue) and emission (red) spectra generated for a.) NR in ethanol, b.) NR in acetonitrile and c.) NR in cyclohexane. The top spectra in each subplot were generated experimentally. The middle spectra in each subplot were obtained by plotting vertical excitations predicted using model set 2, with an SH-EG used calculator to directly infer energy gaps. The bottom spectra in each subplot were generated by passing those same energy gaps into MolSpecPy, then shifting the resulting plots in energy space to have the same peak maxima as the middle spectra.
• Figure S1: Absorption and emission spectra generated for nile red in ethanol by evaluating solute-in-solvent clusters with various radii using four different methods: (a) model set 1, subtracting SH-GS from SH-ES1 energies; (b) model set 2, subtracting SH-GS energies from SH-ES1 energies; (c) model set 2, evaluating clusters with the SH-EG model; and (d) model set 3, evaluating clusters with the MH-EG head.
• Figure S2: Vertical excitation spectra predicted for 5.0 Å radii NR in ethanol clusters using each of the 5 committee members (A-E) of model sets 1 and 2 with the method of subtracting SH-GS from SH-ES1, and also using model set 2 with the SH-EG models to directly infer energy gaps.