arXiv subjects

Milica Todorović

Publications and source records attributed to Milica Todorović.

17 recordsLinked to original sources

arXiv2026

Perspective: Towards sustainable exploration of chemical spaces with machine learning

Artificial intelligence is transforming molecular and materials science, but its growing computational and data demands raise critical sustainability challenges. In this Perspective, we examine resource considerations across the AI-driven discovery pipeline--from quantum-mechanical (QM) data generation and model training to automated, self-driving research workflows--building on discussions from the ``SusML workshop: Towards sustainable exploration of chemical spaces with machine learning'' held in Dresden, Germany. In this context, the availability of large quantum datasets has enabled rigorous benchmarking and rapid methodological progress, while also incurring substantial energy and infrastructure costs. We highlight emerging strategies to enhance efficiency, including general-purpose machine learning (ML) models, multi-fidelity approaches, model distillation, and active learning. Moreover, incorporating physics-based constraints within hierarchical workflows, where fast ML surrogates are applied broadly and high-accuracy QM methods are used selectively, can further optimize resource use without compromising reliability. Equally important is bridging the gap between idealized computational predictions and real-world conditions by accounting for synthesizability and multi-objective design criteria, which is essential for practical impact. Finally, we argue that sustainable progress will rely on open data and models, reusable workflows, and domain-specific AI systems that maximize scientific value per unit of computation, enabling efficient and responsible discovery of technological materials and therapeutics.

cs.LG↗

arXiv2026

Journal Research Data Policies in Materials Science

Open and reproducible research in materials science relies on the availability of data, code, and common metadata standards. Journal research data policies (RDPs) remain a primary mechanism by which publication norms are defined and enforced. We survey RDPs for 171 materials science journals spanning 17 publishers, using an expanded coding framework that captures both data-and-code sharing behavior as well as refereeing standards. We find clear signs of progress in comparison to earlier research on RDPs: nearly all journals provide an RDP, and most mention data availability statements. However, enforceable requirements remain uncommon, public deposition of underlying data is rarely mandatory, and FAIR publication is typically encouraged rather than required. Expectations for research software are substantially less developed than those for data, with limited attention to versioning and persistent identifiers, dependency disclosure, reproducible execution environments, or software quality practices. Aggregating the findings on policy features into an open research data score reveals pronounced heterogeneity across journals. Neither impact factor nor access model reliably predicts policy strength. Double-coding further shows that more complex policies and stricter policies can be more challenging to interpret consistently, and we highlight challenges in consistent RDP encoding across studies. Lastly, we conclude with recommended best practice directions for the future.

cs.DL↗

arXiv2024

Exploring Noncollinear Magnetic Energy Landscapes with Bayesian Optimization

The investigation of magnetic energy landscapes and the search for ground states of magnetic materials using ab initio methods like density functional theory (DFT) is a challenging task. Complex interactions, such as superexchange and spin-orbit coupling, make these calculations computationally expensive and often lead to non-trivial energy landscapes. Consequently, a comprehensive and systematic investigation of large magnetic configuration spaces is often impractical. We approach this problem by utilizing Bayesian Optimization, an active machine learning scheme that has proven to be efficient in modeling unknown functions and finding global minima. Using this approach we can obtain the magnetic contribution to the energy as a function of one or more spin canting angles with relatively small numbers of DFT calculations. To assess the capabilities and the efficiency of the approach we investigate the noncollinear magnetic energy landscapes of selected materials containing 3d, 5d and 5f magnetic ions: Ba$_3$MnNb$_2$O$_9$, LaMn$_2$Si$_2$, $β$-MnO$_2$, Sr$_2$IrO$_4$, UO$_2$ and Ba$_2$NaOsO$_6$. By comparing our results to previous ab initio studies that followed more conventional approaches, we observe significant improvements in efficiency.

cond-mat.mtrl-sci↗

arXiv2024

Question Answering models for information extraction from perovskite materials science literature

Scientific text is a promising source of data in materials science, with ongoing research into utilising textual data for materials discovery. In this study, we developed and tested a novel approach to extract material-property relationships from scientific publications using the Question Answering (QA) method. QA performance was evaluated for information extraction of perovskite bandgaps based on a human query. We observed considerable variation in results with five different large language models fine-tuned for the QA task. Best extraction accuracy was achieved with the QA MatBERT and F1-scores improved on the current state-of-the-art. This work demonstrates the QA workflow and paves the way towards further applications. The simplicity, versatility and accuracy of the QA approach all point to its considerable potential for text-driven discoveries in materials research.

cond-mat.mtrl-sci↗

arXiv2024

Active Learning of Molecular Data for Task-Specific Objectives

Active learning (AL) has shown promise for being a particularly data-efficient machine learning approach. Yet, its performance depends on the application and it is not clear when AL practitioners can expect computational savings. Here, we carry out a systematic AL performance assessment for three diverse molecular datasets and two common scientific tasks: compiling compact, informative datasets and targeted molecular searches. We implemented AL with Gaussian processes (GP) and used the many-body tensor as molecular representation. For the first task, we tested different data acquisition strategies, batch sizes and GP noise settings. AL was insensitive to the acquisition batch size and we observed the best AL performance for the acquisition strategy that combines uncertainty reduction with clustering to promote diversity. However, for optimal GP noise settings, AL did not outperform randomized selection of data points. Conversely, for targeted searches, AL outperformed random sampling and achieved data savings up to 64%. Our analysis provides insight into this task-specific performance difference in terms of target distributions and data collection strategies. We established that the performance of AL depends on the relative distribution of the target molecules in comparison to the total dataset distribution, with the largest computational savings achieved when their overlap is minimal.

cs.LG↗

arXiv2023

Updates to the DScribe Library: New Descriptors and Derivatives

We present an update of the DScribe package, a Python library for atomistic descriptors. The update extends DScribe's descriptor selection with the Valle-Oganov materials fingerprint and provides descriptor derivatives to enable more advanced machine learning tasks, such as force prediction and structure optimization. For all descriptors, numeric derivatives are now available in DSribe. For the many-body tensor representation (MBTR) and the Smooth Overlap of Atomic Positions (SOAP), we have also implemented analytic derivatives. We demonstrate the effectiveness of the descriptor derivatives for machine learning models of Cu clusters and perovskite alloys.

cond-mat.mtrl-sci↗

arXiv2023

Structural disorder by octahedral tilting in inorganic halide perovskites: New insight with Bayesian optimization

Structural disorder is common in metal-halide perovskites and important for understanding the functional properties of these materials. First-principles methods can address structure variation on the atomistic scale, but they are often limited by the lack of structure-sampling schemes required to characterize the disorder. In this work, structural disorder in the benchmark inorganic halide perovskites CsPbI$_3^{}$ and CsPbBr$_3^{}$ is computationally studied in terms of the three octahedral-tilting angles. The consequent variation in energetics and properties are described by three-dimensional potential-energy surfaces (PESs) and property landscapes, delivered by Bayesian Optimization Structure Search method with integrated density-functional-theory (DFT) calculations. The rapid convergence of the PES with about 200 DFT data points in three-dimensional searches demonstrates the power of active learning and strategic sampling with Bayesian optimization. Further analysis indicates that disorder grows with increasing temperature, and reveals that the materials band gap at finite temperatures is a statistical mean over disordered structures.

cond-mat.mtrl-sci↗

arXiv2022

Molecular conformer search with low-energy latent space

Identifying low-energy conformers with quantum mechanical accuracy for molecules with many degrees of freedom is challenging. In this work, we use the molecular dihedral angles as features and explore the possibility of performing molecular conformer search in a latent space with a generative model named variational auto-encoder (VAE). We bias the VAE towards low-energy molecular configurations to generate more informative data. In this way, we can effectively build a reliable energy model for the low-energy potential energy surface. After the energy model has been built, we extract local-minimum conformations and refine them with structure optimization. We have tested and benchmarked our low-energy latent-space (LOLS) structure search method on organic molecules with $5-9$ searching dimensions. Our results agree with previous studies.

physics.comp-ph↗

arXiv2021

Integrating Bayesian Inference with Scanning Probe Experiments for Robust Identification of Surface Adsorbate Configurations

Controlling the properties of organic/inorganic materials requires detailed knowledge of their molecular adsorption geometries. This is often unattainable, even with current state-of-the-art tools. Visualizing the structure of complex non-planar adsorbates with atomic force microscopy (AFM) is challenging, and identifying it computationally is intractable with conventional structure search. In this fresh approach, cross-disciplinary tools are integrated for a robust and automated identification of 3D adsorbate configurations. Bayesian optimization is employed with first-principles simulations for accurate and unbiased structure inference of multiple adsorbates. The corresponding AFM simulations then allow fingerprinting adsorbate structures that appear in AFM experimental images. In the instance of bulky (1S)-camphor adsorbed on the Cu(111) surface, three matching AFM image contrasts are found, which allow correlating experimental image features to distinct cases of molecular adsorption.

cond-mat.mtrl-sci↗

arXiv2020

Detecting stable adsorbates of (1S)-camphor on Cu(111) with Bayesian optimization

Identifying the atomic structure of organic-inorganic interfaces is challenging with our current research tools. Interpreting the structure of complex molecular adsorbates from microscopy images can be difficult, and using atomistic simulations to find the most stable structures is limited to partial exploration of the potential energy surface due to the high-dimensional phase space. In this study, we present the recently developed Bayesian Optimization Structure Search (BOSS) method as an efficient solution for identifying the structure of non-planar adsorbates. We apply BOSS with density-functional theory simulations to detect the stable adsorbate structures of (1S)-camphor on the Cu(111) surface. We identify the optimal structure among 8 unique types of stable adsorbates, in which camphor chemisorbs via oxygen (global minimum) or physisorbs via hydrocarbons to the Cu(111) surface. This study demonstrates that new cross-disciplinary tools, like BOSS, facilitate the description of complex surface structures and their properties, and ultimately allow us to tune the functionality of advanced materials.

cond-mat.mtrl-sci↗

arXiv2020

Predicting Gas-Particle Partitioning Coefficients of Atmospheric Molecules with Machine Learning

The formation, properties and lifetime of secondary organic aerosols in the atmosphere are largely determined by gas-particle partitioning coefficients of the participating organic vapours. Since these coefficients are often difficult to measure or compute, we developed a machine learning (ML) model to predict them given molecular structure as input. Our data-driven approach is based on the dataset by Wang et al. (Atmos. Chem. Phys., 17, 7529 (2017)), who computed the partitioning coefficients and saturation vapour pressures of 3414 atmospheric oxidation products from the master chemical mechanism using the COSMOtherm program. We train a kernel ridge regression (KRR) ML model on the saturation vapour pressure ($P_{sat}$), and on two equilibrium partitioning coefficients: between a water-insoluble organic matter phase and the gas phase ($K_{WIOM/G}$), and between an infinitely dilute solution with pure water and the gas phase ($K_{W/G}$). For the input representation of the atomic structure of each organic molecule to the machine, we test different descriptors. Our best ML model predicts $P_{sat}$ and $K_{WIOM/G}$ to within 0.3 and $K_{W/G}$ to within 0.4 logarithmic units of the original COSMOtherm calculations. This is equal or better than the typical accuracy of COSMOtherm predictions compared to experimental data. We then apply our ML model to a dataset of 35,383 molecules that we generated based on a carbon 10 backbone and functionalized with 0 to 6 carboxyl, carbonyl or hydroxyl groups to evaluate its performance for polyfunctional compounds with potentially low $P_{sat}$. The resulting $P_{sat}$ and partitioning coefficient distributions were physico-chemically reasonable, and the volatility predictions for the most highly oxidized compounds were in qualitative agreement with experimentally inferred volatilities of atmospheric oxidation products with similar elemental composition.

physics.chem-ph↗

arXiv2020

Projective Preferential Bayesian Optimization

Bayesian optimization is an effective method for finding extrema of a black-box function. We propose a new type of Bayesian optimization for learning user preferences in high-dimensional spaces. The central assumption is that the underlying objective function cannot be evaluated directly, but instead a minimizer along a projection can be queried, which we call a projective preferential query. The form of the query allows for feedback that is natural for a human to give, and which enables interaction. This is demonstrated in a user experiment in which the user feedback comes in the form of optimal position and orientation of a molecule adsorbing to a surface. We demonstrate that our framework is able to find a global minimum of a high-dimensional black-box function, which is an infeasible task for existing preferential Bayesian optimization frameworks that are based on pairwise comparisons.

stat.ML↗

arXiv2020

Efficient hyperparameter tuning for kernel ridge regression with Bayesian optimization

Machine learning methods usually depend on internal parameters -- so called hyperparameters -- that need to be optimized for best performance. Such optimization poses a burden on machine learning practitioners, requiring expert knowledge, intuition or computationally demanding brute-force parameter searches. We here address the need for more efficient, automated hyperparameter selection with Bayesian optimization. We apply this technique to the kernel ridge regression machine learning method for two different descriptors for the atomic structure of organic molecules, one of which introduces its own set of hyperparameters to the method. We identify optimal hyperparameter configurations and infer entire prediction error landscapes in hyperparameter space, that serve as visual guides for the hyperparameter dependence. We further demonstrate that for an increasing number of hyperparameters, Bayesian optimization becomes significantly more efficient in computational time than an exhaustive grid search -- the current default standard hyperparameter search method -- while delivering an equivalent or even better accuracy.

physics.chem-ph↗

arXiv2020

Atomic structures and orbital energies of 61,489 crystal-forming organic molecules

Data science and machine learning in materials science require large datasets of technologically relevant molecules or materials. Currently, publicly available molecular datasets with realistic molecular geometries and spectral properties are rare. We here supply a diverse benchmark spectroscopy dataset of 61,489 molecules extracted from organic crystals in the Cambridge Structural Database (CSD), denoted OE62. Molecular equilibrium geometries are reported at the Perdew-Burke-Ernzerhof (PBE) level of density functional theory (DFT) including van der Waals corrections for all 62k molecules. For these geometries, OE62 supplies total energies and orbital eigenvalues at the PBE and the PBE hybrid (PBE0) functional level of DFT for all 62k molecules in vacuum as well as at the PBE0 level for a subset of 30,876 molecules in (implicit) water. For 5,239 molecules in vacuum, the dataset provides quasiparticle energies computed with many-body perturbation theory in the $G_0W_0$ approximation with a PBE0 starting point (denoted GW5000 in analogy to the GW100 benchmark set (M. van Setten et al. J. Chem. Theory Comput. 12, 5076 (2016))).

physics.comp-ph↗

arXiv2019

Chemical diversity in molecular orbital energy predictions with kernel ridge regression

Instant machine learning predictions of molecular properties are desirable for materials design, but the predictive power of the methodology is mainly tested on well-known benchmark datasets. Here, we investigate the performance of machine learning with kernel ridge regression (KRR) for the prediction of molecular orbital energies on three large datasets: the standard QM9 small organic molecules set, amino acid and dipeptide conformers, and organic crystal-forming molecules extracted from the Cambridge Structural Database. We focus on prediction of highest occupied molecular orbital (HOMO) energies, computed at density-functional level of theory. Two different representations that encode molecular structure are compared: the Coulomb matrix (CM) and the many-body tensor representation (MBTR). We find that KRR performance depends significantly on the chemistry of the underlying dataset and that the MBTR is superior to the CM, predicting HOMO energies with a mean absolute error as low as 0.09 eV. To demonstrate the power of our machine learning method, we apply our model to structures of 10k previously unseen molecules. We gain instant energy predictions that allow us to identify interesting molecules for future applications.

physics.chem-ph↗

arXiv2019

Efficient Bayesian Inference of Atomistic Structure in Complex Functional Materials

Tailoring the functional properties of advanced organic/inorganic heterogeonous devices to their intended technological applications requires knowledge and control of the microscopic structure inside the device. Atomistic quantum mechanical simulation methods deliver accurate energies and properties for individual configurations, however, finding the most favourable configurations remains computationally prohibitive. We propose a 'building block'-based Bayesian Optimisation Structure Search (BOSS) approach for addressing extended organic/inorganic interface problems and demonstrate its feasibility in a molecular surface adsorption study. In BOSS, a likelihood-free Bayesian scheme accelerates the identification of material energy landscapes with the number of sampled configurations during active learning, enabling structural inference with high chemical accuracy and featuring large simulation cells. This allowed us to identify several most favourable molecular adsorption configurations for $\mathrm{C}_{60}$ on the (101) surface of $\mathrm{TiO}_2$ anatase and clarify the key molecule-surface interactions governing structural assembly. Inferred structures were in good agreement with detailed experimental images of this surface adsorbate, demonstrating good predictive power of BOSS and opening the route towards large-scale surface adsorption studies of molecular aggregates and films.

cond-mat.mtrl-sci↗

arXiv2013

Density-functional theory study of gramicidin A ion channel geometry and electronic properties

Understanding the mechanisms underlying ion channel function from the atomic-scale requires accurate ab initio modelling as well as careful experiments. Here, we present a density functional theory (DFT) study of the ion channel gramicidin A, whose inner pore conducts only monovalent cations and whose conductance has been shown to depend on the side chains of the amino acids in the channel. We investigate the ground-state geometry and electronic properties of the channel in vacuum, focusing on their dependence on the side chains of the amino acids. We find that the side chains affect the ground state geometry, while the electrostatic potential of the pore is independent of the side chains. This study is also in preparation for a full, linear scaling DFT study of gramicidin A in a lipid bilayer with surrounding water. We demonstrate that linear scaling DFT methods can accurately model the system with reasonable computational cost. Linear scaling DFT allows ab initio calculations with 10,000 to 100,000 atoms and beyond, and will be an important new tool for biomolecular simulations.

q-bio.BM↗