Thermoplastic starch (TPS) has emerged as a biodegradable alternative to conventional plastics. However, the literature indicates that its functional performance varies with its composition and processing conditions, hindering efficient material design with traditional trial-and-error approaches. To address this need, this study develops machine learning-based predictive models to estimate the mechanical, thermal, permeability, barrier, and biodegradability properties of TPS as a function of formulation and processing variables. A dataset of 98 variables and 962 observations, obtained from previous studies, was constructed for this purpose. The CRISP-ML(Q) methodology was followed, and three supervised algorithms were implemented: Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Artificial Neural Networks (ANN). Model performance was evaluated using hold-out validation (70:20:10) and standard regression metrics. The results showed that XGBoost is the algorithm with the best overall predictive ability. The highest levels of accuracy were achieved for WVP (R2 = 0.981) and Td (R2 = 0.939). The main predictive limitations were observed in some ANN models, especially with the variable’s Tensile strength, Young's modulus, and T50, which reported negative or near-zero R2 values. The developed models constitute a practical decision-support tool, allowing researchers and non-specialist users to virtually evaluate different combinations of formulation and processing conditions before conducting laboratory tests, thus reducing the number of experiments during the initial stages of new material development.
Recent advances in large language models (LLMs) have opened new avenues for automating scientific research workflows. Systems like the AI Scientist [1] demonstrate fully autonomous endto-end research pipelines for machine learning, but their applicability to domains requiring expensive high-performance computing (HPC) resources, specialized quantum chemistry software, and rigorous thermochemical accuracy remains unexplored. Here we present VirtualLab_CC, a hybrid LLM-augmented virtual laboratory for computational chemistry that integrates Claude (Anthropic) as an orchestration layer with a multi-agent architecture, gated lifecycle management, and bidirectional HPC cluster synchronization. Unlike fully autonomous systems, VirtualLab_CC enforces human-in-the-loop quality gates at every critical decision point while automating routine operations: conformer searches, input generation, job submission and monitoring, output parsing, failure diagnosis, and manuscript preparation. The system manages 21 specialized skills spanning literature review, cluster operations, result extraction, novelty verification, AI-assisted peer review, and citation validation. We demonstrate VirtualLab_CC on a prototype case study of 1,3-dipolar cycloadditions between stabilized Criegee intermediates and biogenic terpene dipolarophiles, involving 17 molecular species, 34 conformers, and over 100 SLURM jobs across multiple calculation stages. The system successfully identified and autocorrected cluster failures, detected a methodological anomaly (a suspiciously low dioxirane barrier requiring IRC verification), and coordinated multi-method validation across three DFT functionals. VirtualLab_CC bridges the gap between fully autonomous AI research and the rigorous quality control demands of computational chemistry, providing a practical framework for LLM-augmented scientific workflows in resource-constrained HPC environments.
Accurate prediction and generation of reaction temperature remain open challenges in data-driven chemical synthesis, owing to the continuous nature heterogeneous experimental protocols, and the need for strict physical plausibility. To address this, we introduce the Coherent Guided Generative Model (CGGM), a hybrid deep-learning architecture that integrates variational inference, adversarial learning, and smooth quadratic physical constraints to map reaction temperatures directly from molecular descriptors. Benchmarked on a rigorously curated dataset of 554,974 reactions from the Open Reaction Database (ORD), the CGGM exposes a fundamental precision-generalization trade-off in chemical AI. While a standalone Variational Autoencoder (VAE) minimizes pointwise error (MAE = 0.5105) through conservative mean-regression, the hybrid CGGM achieves superior global density estimation, yielding the highest log-likelihood (-5.2621) and accurately capturing rare cryogenic and high-temperature regimes. Crucially, the CGGM enforces thermodynamic boundaries, restricting out-of-range predictions to just 0.05%, nearly identical to the curated empirical ground truth (0.04%) and significantly outperforming a standalone GAN (0.16%). Latent space topology and manifold traversal analyses confirm a smooth, continuous embedding space with monotonic temperature encoding (rho > 0.99, p < 0.001), allowing seamless interpolation between cryogenic (approximate to-78 degrees C and reflux (>100 degrees C) domains without structural degradation. Finally, stratified evaluations across Suzuki, Buchwald-Hartwig, and esterification reactions highlight that while classical regressors are optimal for static pointwise lookups, the primary merit of the CGGM lies in its exceptional distributional fidelity, tail-risk modeling, and physically bounded generation. This hybrid framework establishes a robust, uncertainty-aware computational engine for continuous chemical space exploration and closed-loop autonomous synthesis planning.
Quantitative simulation of trivalent f-block chelates in water remains challenging because bonded and non-bonded force-field models make different approximations for coordination structure, exchange dynamics, and ion–ligand interactions in highly charged systems. Here, we develop a hybrid machine-learning/molecular-mechanics (ML/MM) framework for Ac3+–DOTA in explicit solvent by training an E(3)-equivariant neural network potential (MACELES) on mechanically embedded QM/MM data for Ac aquo and Ac–DOTA species and coupling it to NAMD 2.14 with particle-mesh Ewald electrostatics. Nanosecond ML/MM trajectories remain numerically stable and preserve chelate integrity, yielding a compact DOTA inner shell with an inner-sphere water coordination number of CNAc,Ow≈1.7 arising from a dynamic equilibrium between one- and two-water states (37.5% and 59.9% of frames; three waters 2.5%). A 5 ns potential of mean force shows two low-lying basins at CNAc,Ow≈1 and CNAc,Ow≈2. DFT end-state free energies are consistent with the ML/MM profile, and DFT minimum-energy paths provide a qualitative electronic-structure reference for the observed basin connectivity. State-resolved kinetics reveal picosecond water-exchange pathways that couple hydration changes to transient DOTA arm fluctuations, and training-set comparisons show that temperature-matched Ac–DOTA data optimize energy/force accuracy while more diverse solvated data improve charge prediction. Overall, the present hybrid ML/MM model provides a practical description of Ac3+–DOTA hydration thermodynamics and short-time exchange behavior in explicit water at MD-like cost.
Identifying explosive molecules is crucial for safety and materials research. In this work, we propose a machine learning framework for classifying 119 molecules using images of their electrostatic potential maps (MEPs), computed from quantum-chemical methods and depicting the molecules' positive and negative electric charge regions. The training dataset was augmented by rotating each image 30 times, and a Convolutional Neural Network (CNN) was subsequently trained to classify molecules. The CNN classification model was combined with advanced gradient-based Explainable Artificial Intelligence (XAI) techniques and visualization methods. The CNN achieved 100% accuracy on the test set, demonstrating its effectiveness. In a separate experiment, it achieved a 81.25% accuracy using molecules chemically distinct from the training set: the CNN model classified eight as explosive, including two known simulants and two non-explosive compounds containing strong electron-withdrawing groups, thus indicating sensitivity to chemically relevant features not included in the training distribution. To gain a deeper understanding of the model’s decision-making process, we used the Saliency Map technique with the reduce_max function to identify the most critical regions in the MEP images for the classification task. Neutral and homogeneous MEP regions primarily characterize non-explosive molecules. In contrast, explosive compounds are associated with pronounced electrostatic potential gradients and localized charge features, particularly in regions related to nitro groups and trigger bonds. This work highlights the potential of combining MEPs computed from quantum-chemical methods with machine learning to analyze and classify molecules with targeted properties.
