Rubens Caio Souza

Accurate reference data for electronically excited states are essential for assessing multireference electronic-structure methods. We benchmark state-averaged completeactive-space self-consistent field (SA-CASSCF), exchange-scaled CASSCF (XS-CASSCF), fully uncontracted multireference configuration interaction with single and double excitations (MR-CISD), and four a posteriori size-consistency corrections against the theoretical best estimates of QUESTDB. The dataset comprises 262 valence excitations of 41 small and medium-sized molecules, including singlet and triplet n*, *, and doubly excited states. All calculations used a consistent COLUMBUS protocol with molecule-specific active spaces and the aug-cc-pVDZ basis set. For the 232 safe transitions, SA-CASSCF and MR-CISD yield mean absolute errors of 0.57 and 0.31 eV, respectively, with predominantly positive biases. The Davidson–Silver/Siegbahn (+DV3) and Pople (+P) corrections reduce the mean absolute errors to 0.10 and 0.11 eV and the root-mean-square errors to 0.13 eV. They also provide a more balanced description across spin multiplicities and excitation characters. Both XSCASSCF improve singlet * states, particularly those with pronounced ionic character, but degrade n* states and do not improve the overall statistics relative to SA-CASSCF. For 36 selected unsafe or doubly excited transitions, MR-CISD+DV3 and MR-CISD+P retain mean absolute errors of 0.16 eV. Their agreement with exFCI-based reference values and discrepancies from several CASPT3- and PC-NEVPT2-based estimates identify states whose current best estimates merit independent re-examination. Overall, size-consistencycorrected MR-CISD provides highly accurate and broadly balanced valence excitation energies.

A small language model (SLM)-based workflow was developed to analyze over 5,000 Brazilian Symposium of Theoretical Chemistry (SBQT) one-page abstracts (1981-2023). Brazilian theoretical and computational chemistry expanded strongly until the 2000s, then declined and stabilized in the 2010s and early 2020s. The number of abstracts and participants has stabilized at around 300 since 1997. Research funding has been dominated by the Brazilian agencies National Council for Scientific and Technological Development (CNPq) and Agency for Support and Evaluation of Graduate Education (CAPES), but state foundations have made relevant contributions. The most popular theoretical methods over time include configuration interaction (CI), molecular dynamics (MD), and density functional theory (DFT). The initial applications of DFT followed the rest of the world, but investigations using the most advanced exchange-correlation functionals and dispersion corrections lag by 5-8 years in some groups. Local development of methods and software is very incipient. It was identified significant delays in Brazilian theoretical and computational chemistry, relative to the rest of the world, in machine learning techniques, automated workflows, high-performance graphics processing unit (GPU) architectures, and benchmarking. As most studies focus on molecular electronic structure calculations, the most widely used software is for this type of research. The most studied subjects are intra- and inter‑proton, electron, and charge-transfer chemical phenomena. Hydrogen bonding and van der Waals interactions have also been widely studied, with water and benzene being the most investigated molecules. These findings highlight possible pathways for Brazilian theoretical chemistry to increase its reach. In particular, data and code repositories that enable reproducible workflows, which are largely absent, would increase reach and comply with the highly desirable findable, accessible, interoperable, and reusable (FAIR) practices.

Accurately predicting thermophysical properties across various physical states is essential for both industrial and scientific applications. However, experimental data often exhibit variability and noise, requiring robust modeling approaches. In this work, we employ machine learning (ML) techniques to predict methane's thermophysical properties in liquid, vapor, and supercritical phases, including isobaric and isochoric heat capacities, density, specific volume, Joule-Thomson coefficients, enthalpies, sound speed, and shear viscosities, applying an approach recently developed [ACS Eng. Au 5, 226 (2025)10.1021/acsengineeringau.5c00001]. We explore various ML algorithms, including Adaptive Boosting, Bagging, Decision Trees, Extra Trees, Gradient Boosting, Histogram-based Gradient Boosting Regression Tree, K-Nearest Neighbors, Light Gradient Boosting Machine, Nu-Support Vector Regression, Random Forest, Extreme Gradient Boosting, and Artificial Neural Networks. The ML models, which use raw experimental data without removing outliers, produce predictions that are closer to those obtained from the equations of state (EOS) employed by the National Institute of Standards and Technology than to the same raw experimental data used as input for developing the EOS. These results highlight ML's potential to identify and generalize complex patterns, smooth inherent noise, and manage the variability of different thermophysical properties. They indicate that ML models, particularly Extra Trees and Gradient Boosting, can offer a scalable alternative for thermophysical property predictions, offering consistency and efficiency over traditional methods. Although our approach does not eliminate preprocessing, it demonstrates that ML can effectively manage noisy data independently, offering a more efficient and cost-effective alternative to conventional pre- and postprocessing techniques.

Large Language Models (LLMs) are a machine learning technique that has transformed natural language processing. However, their large computational demands limit their accessibility, leading to the development of Small Language Models (SLMs), which, by running locally on a microcomputer, made AI-driven language processing and enhanced control for text analysis widely accessible. In this work, we use an SLM to analyze the evolution of Chemistry in Brazil by comparing data from the 2019 and 2024 Brazilian Chemical Society meetings (RASBQ). We demonstrate the viability of SLMs for extracting and structuring large volumes of text from scientific events collected in books of abstracts, thus enabling comprehensive comparative analyses that would otherwise be impractical. Our methodology extracts abstracts from the RASBQ digital proceedings and processes them using SLMs. These models converted the textual content into structured, manipulable data, thus enabling us to conduct a semantic and statistical analysis of the two events. The results highlight how SLMs can efficiently transform unstructured scientific proceedings into tractable data, thereby saving significant time and resources. The comparison between the 2019 and 2024 events revealed notable changes in thematic distribution, institutional participation, and potential regional impacts, underscoring the importance of data standardization in automated analyses. A Portuguese version of the Results and Discussion section is presented in the Supplementary Information.

The search for functional fluorescent organic materials can significantly benefit from rapid and accurate predictions of photophysical properties. However, screening large numbers of potential fluorophore molecules in different solvents faces the limitations of quantum mechanical calculations and experimental measurements. In this work, we develop machine learning (ML) algorithms for predicting the fluorescence of a molecule, focusing on two target properties: emission wavelengths (WLs) and quantum yields (QYs). For this purpose, we employ the Deep4Chem database containing optical properties of 20,236 combinations of 7,016 chromophores in 365 different solvents. Several chemical descriptors, or features, were selected as inputs for each model, and each molecule was characterized by its SMILES fingerprint. The Shapley Additive explanations (SHAP) technique was used to rationalize the results, showing that the most impactful properties are chromophore-related, as expected from chemical intuition. For the best-performing model, the Random Forest, our results for the test set show a root-mean-square error (RMSE) of 28.8 nm (0.15 eV) for WLs and 0.19 for QYs. The developed ML models were used to predict, thus completing, the missing results for the WL and QY target properties in the original Deep4Chem database, resulting in two new databases, one for each property. Testing our ML models for each target property in molecules not included in the original Deep4Chem database presented good results.

Large language models (LLMs) are a machine learning technique that has transformed natural language processing. However, their large computational demands limit their accessibility, leading to the development of small language models (SLMs), which, by running locally on a microcomputer, made artificial intelligence (AI)-driven language processing and enhanced control for text analysis widely accessible. In this work, we use an SLM to analyze the evolution of Chemistry in Brazil by comparing data from the 2019 and 2024 Brazilian Chemical Society meetings (RASBQ). We demonstrate the viability of SLMs for extracting and structuring large volumes of text from scientific events collected in books of abstracts, thus enabling comprehensive comparative analyses that would otherwise be impractical. Our methodology extracts abstracts from the RASBQ digital proceedings and processes them using SLMs. These models converted the textual content into structured, manipulable data, thus enabling us to conduct a semantic and statistical analysis of the two events. The results highlight how SLMs can efficiently transform unstructured scientific proceedings into tractable data, thereby saving significant time and resources. The comparison between the 2019 and 2024 events revealed notable changes in thematic distribution, institutional participation, and potential regional impacts, underscoring the importance of data standardization in automated analyses.

Theoretical and computational chemistry (TCC) is a set of theories and models that, over the years, were refined to the point that it is possible to determine measurable quantities with precision, predict experimental results, and provide fundamental insights into chemical phenomena and mechanisms that may be difficult or impossible to observe experimentally. Machine Learning (ML), on the other hand, is a subfield of Artificial Intelligence (AI) that applies different types of statistical methods to a large volume of data or a smaller volume of precise data, combined with high computational power, enabling the discovery of complex patterns and production of explanations inaccessible through human deductive reasoning and intuition alone or using traditional scientific methods. Recently, ML, combined or not with TCC methods, has emerged as a transformative force, bringing significant advances in chemistry and materials science. This review surveys basic ML concepts and their applications in chemistry, focusing on supervised and unsupervised learning approaches, data preprocessing, and model development workflows, exploring the most relevant ML algorithms selected for their specific usefulness in chemical applications. Integrating ML with traditional computational chemistry methods, such as density functional theory, is highlighted as a powerful synergy for accelerating materials discovery and design. Key areas of impact discussed include High-Throughput Virtual Screening (HTVS) of molecules and materials, spectroscopy (including UV-Vis and fluorescence), organic electronics (such as solar cells and organic light-emitting diodes), potential energy surfaces, and molecular dynamics. It also addresses critical aspects of ML in chemistry, including data representation, model interpretability through explainability techniques, and the emerging role of large language models (LLM). Tips for acquiring knowledge of ML for practical applications in chemistry are given.

Accurate knowledge of electronic molecular properties of excited states is fundamental for understanding the behavior of functional materials for organic electronics and sensors. In this work, we focus on determining the properties of the most intense peak in the electronic absorption spectra of organic molecules. For this purpose, we employed the quantum chemistry QM-symex dataset, which has approximately 173,000 organic molecules and time-dependent density functional theory (TD-DFT) data of the first ten electronic absorption transitions. Each one is identified by its Cartesian coordinates. From data in the original QM-symex, we built a new dataset named QM-symex-modif that contains molecules in simplified molecular input line entry system (SMILES) format and properties related to the main electronic transition. We then employed twenty machine learning (ML) algorithms to investigate oscillator strengths, excitation energies, transition orbitals, and the highest occupied molecular orbitals (HOMOs). As inputs for the ML algorithms, we used several chemical descriptors for each molecule generated in the RDKit tool employing the corresponding SMILES format. The generated input descriptors significantly improved the accuracy of the ML predictions for these key photophysical properties. Very good mean absolute errors (MAEs) were obtained for the test set composed of 45,056 molecules, namely, an MAE of 0.035 for oscillator strengths, 0.09 eV for excitation energies, 1.24 and 0.62 for the initial and final transition molecular orbital (MO) numbers (i.e., for each molecule, their position in the MO listing), respectively, and 0.014 for HOMO numbers, with coefficient of determination (R2 ) values consistently exceeding 0.94, thus demonstrating the accuracy of the models. Additionally, a Shapley additive explanation (SHAP) analysis was carried out to evaluate the importance of the input parameters for the investigated ML models. We found several interesting relationships involving the input parameters. In particular, molecular weight holds significant importance in our ML models for determining the target HOMO numbers and the transition orbitals.

Accurate knowledge of electronic molecular properties of excited states is fundamental for understanding the behavior of functional materials for organic electronics and sensors. In this work, we focus on determining the properties of the most intense peak in the electronic absorption spectra of organic molecules. For this purpose, we employed the quantum chemistry QM-symex dataset, which has approximately 173,000 organic molecules and time-dependent DFT (TD-DFT) data of the first ten electronic absorption transitions. Each one is identified by its Cartesian coordinates. From data in the original QM-symex, we built a new dataset named QM-symex-modif that contains molecules in Simplified Molecular Input Line Entry System (SMILES) format and properties related to the main electronic transition. We then employed twenty machine learning (ML) algorithms to investigate oscillator strengths, excitation energies, transition orbitals, and the highest occupied molecular orbitals (HOMOs). As inputs for the ML algorithms, we used several chemical descriptors for each molecule generated in the RDKit tool employing the corresponding SMILES format. The generated input descriptors significantly improved the accuracy of the ML predictions for these key photophysical properties. Very good mean absolute errors (MAEs) were obtained for the test set composed of 45,056 molecules, namely, an MAE of 0.035 for oscillator strengths, 0.09 eV for excitation energies, 1.24 and 0.62 for the initial and final transition molecular orbital (MO) numbers (i.e., for each molecule, their position in the MO listing) respectively, and 0.014 for HOMO numbers, with R² values consistently exceeding 0.94, thus demonstrating the accuracy of the models. Additionally, a Shapley additive explanation (SHAP) analysis was carried out to evaluate the importance of the input parameters for the investigated ML models. We found several interesting relationships involving the input parameters. In particular, molecular weight holds significant importance in our ML models for determining the target HOMO numbers and the transition orbitals.