Max Pinheiro

Mixed quantum–classical dynamics (MQCD) methods are effective models for excited-state processes in quasi-classical molecular systems, in which nuclear motion is described by classical trajectories while electronic populations undergo quantum nonadiabatic transitions. This article presents Newton-X 26, a new generation of the Newton-X platform that consolidates two decades of development into a modular ecosystem for generating spectra and initial conditions, propagating dynamics, and analyzing, postprocessing, and archiving data. Newton-X 26 supports multiple MQCD strategies, including surface hopping, decoherence-corrected Ehrenfest dynamics, and ab initio multiple spawning, and connects to a range of electronic-structure engines through dedicated interfaces. The platform emphasizes efficient execution for large trajectory ensembles, enabling systematic convergence analyses and uncertainty estimation. Complementary tools support automated data curation, machine-learning-assisted workflows, and reproducible FAIR-oriented reporting and sharing. Taken together, Newton-X 26 provides an open-source environment for routine MQCD applications and continued method development across multiple electronic-structure levels.

Machine learning (ML) models are increasingly used in quantum chemistry, but their reliability hinges on uncertainty quantification (UQ). In this study, we compare two prominent UQ paradigms—deep evidential regression (DER) and deep ensembles—on the QM9 and WS22 datasets, with a specific emphasis on the role of post hoc calibration. Raw uncertainties from both methods were systematically miscalibrated: DER produced uncertainty estimates where data noise and model uncertainty were not cleanly separated, while ensembles produced sharper yet underconfident estimates. Applying calibration techniques such as isotonic regression (ISR), standard scaling, and GP‐Normal corrected these deficiencies, aligning predicted variances with observed errors. On QM9, calibration enabled DER to filter high‐confidence predictions more effectively than ensembles. On WS22, calibrated ensembles not only improved statistical reliability but also delivered substantial computational savings in active learning, reducing redundant ab initio evaluations by more than 20%. These results demonstrate that post hoc calibration is essential to transform uncertainty estimates from descriptive metrics into actionable signals, ensuring both trustworthy predictions and resource‐efficient molecular modeling.

Machine learning (ML) models are increasingly used in quantum chemistry, but their reliability hinges on uncertainty quantification (UQ). In this study, we compare two prominent UQ paradigms—Deep Evidential Regression (DER) and Deep Ensembles—on the QM9 and WS22 datasets, with a specific emphasis on the role of post hoc calibration. Raw uncertainties from both methods were systematically miscalibrated: DER produced uncertainty estimates where data noise and model uncertainty were not cleanly separated, while ensembles produced sharper yet underconfident estimates. Applying calibration techniques such as isotonic regression, standard scaling, and GP-Normal corrected these deficiencies, aligning predicted variances with observed errors. On QM9, calibration enabled DER to filter high-confidence predictions more effectively than ensembles. On WS22, calibrated ensembles not only improved statistical reliability but also delivered substantial computational savings in active learning, reducing redundant ab initio evaluations by more than 20%. These results demonstrate that post hoc calibration is essential to transform uncertainty estimates from descriptive metrics into actionable signals, ensuring both trustworthy predictions and resource-efficient molecular modeling.

Exploring molecular excited states holds immense significance across organic chemistry, chemical biology, and materials science.

Photochemical processes span timescales from femtoseconds to nanoseconds, and their simulation via Fewest Switches Surface Hopping (FSSH) requires a large number of computationally expensive electronic structure evaluations. Machine learning (ML) interatomic potentials can reduce this cost, but they must be trained on datasets that capture the most relevant regions of configurational space. We present MELTS, a fully automated active learning (AL) program for FSSH that iteratively improves ML models by using trajectory propagation to guide sampling. MELTS integrates Newton-X and MLatom through socket-based communication, minimizing I/O overhead and enabling large- scale simulations with a user-friendly interface. We validate the AL protocol implemented in MELTS on two contrasting systems: ultrafast fulvene dynamics (tens of femtoseconds) and nanosecond-scale pyrene fluorescence. In both cases, MELTS delivers quantitative agreement with reference quantum results while reducing computational time by up to three orders of magnitude. This demonstrates that MELTS can efficiently generate accurate ML potentials for photochemical processes across a wide range of timescales.

ULaMDyn automates the unsupervised analysis of high-dimensional nonadiabatic molecular dynamics data, enabling the discovery of key properties critical for excited-state dynamics.

Polycyclic aromatic hydrocarbons (PAHs) exhibit intriguing characteristics that position them as promising candidates for advancements in organic semiconductor technologies. Notably, tetracene finds substantial utility in Electronics due to its application in organic light-emitting diodes (OLEDs) and organic field-effect transistors (OFETs). The strategic introduction of an isoelectronic boron-nitrogen (B,N) pair to replace a carbon-carbon pair in acenes introduces changes in the electronic structure, allowing for the controlled modulation of diradical characteristics. Consequently, this B,N substitution enables precise adjustments in chemical, optical, and electronic attributes. In this work, we undertook a systematic exploration of thermally activated delayed fluorescence (TADF) phenomena within a set of 77 B,N-substituted derivatives of tetracene. The primary objective was to identify and select prospective molecules for the fabrication of OLEDs. Employing multiconfigurational methods of computational quantum chemistry, we conducted an extensive investigation to unravel the potential candidates. As a result, we identified molecules that might exhibit the sought-after TADF behavior. Descriptors such as excitation energies, harmonic oscillator model of aromaticity (HOMA) and fractional occupation number weighted density (FOD) were assessed and indicated five candidates with stability comparable to that of pristine tetracene. This research not only contributes to a deeper understanding of the influence of B,N substitution on acene derivatives but also opens doors for the development of organic electronics by harnessing the properties of these selected molecules.

The analysis of nonadiabatic molecular dynamics (NAMD) data presents significant challenges due to its high dimensionality and complexity. To address these issues, we introduce ULaMDyn, a Python-based, open-source package designed to automate the unsupervised analysis of large datasets generated by NAMD simulations. ULaMDyn integrates seamlessly with the Newton-X platform and employs advanced dimensionality reduction and clustering techniques to uncover hidden patterns in molecular trajectories, enabling a more intuitive understanding of excited-state processes. Using the photochemical dynamics of fulvene as a test case, we demonstrate how ULaMDyn efficiently identifies critical molecular geometries and critical nonadiabatic transitions. The package offers a streamlined, scalable solution for interpreting large NAMD datasets. It is poised to facilitate advances in the study of excited-state dynamics across a wide range of molecular systems.

Machine learning (ML) is increasingly becoming a common tool in computational chemistry. At the same time, the rapid development of ML methods requires a flexible software framework for designing custom workflows. MLatom 3 is a program package designed to leverage the power of ML to enhance typical computational chemistry simulations and to create complex workflows. This open-source package provides plenty of choice to the users who can run simulations with the command-line options, input files, or with scripts using MLatom as a Python package, both on their computers and on the online XACS cloud computing service at XACScloud.com. Computational chemists can calculate energies and thermochemical properties, optimize geometries, run molecular and quantum dynamics, and simulate (ro)vibrational, one-photon UV/vis absorption, and two-photon absorption spectra with ML, quantum mechanical, and combined models. The users can choose from an extensive library of methods containing pretrained ML models and quantum mechanical approximations such as AIQM1 approaching coupled-cluster accuracy. The developers can build their own models using various ML algorithms. The great flexibility of MLatom is largely due to the extensive use of the interfaces to many state-of-the-art software packages and libraries.

This study investigates velocity adjustment directions after hopping in surface hopping dynamics. Using fulvene and a protonated Schiff base (PSB4) as case studies, we investigate the population decay and reaction yields of different sets of dynamics with the velocity adjusted in either the nonadiabatic coupling, gradient difference, or momentum directions. For the latter, in addition to the conventional algorithm, we investigated the performance of a reduced kinetic energy reservoir approach recently proposed. Our evaluation also considered velocity adjustment in the directions of approximate nonadiabatic coupling vectors. While results for fulvene are susceptible to the adjustment approach, PSB4 is not. We correlated this dependence to the topography near the conical intersections. When nonadiabatic coupling vectors are unavailable, the gradient difference direction is the best adjustment option. If the gradient difference is also unavailable, a semiempirical vector direction or the momentum direction with a reduced kinetic energy reservoir becomes an excellent option to prevent an artificial excess of back hoppings. The precise velocity adjustment direction is less crucial for describing the nonadiabatic dynamics than the kinetic energy reservoir's size.