The development of artificial intelligence, and in particular machine learning, has seen tremendous success in recent years. The use of machine learning, however, extends to vast application areas outside of those that we encounter in our day-to-day life. One such area is within the natural sciences, where research has shown promising results in using machine learning for modeling of systems of atoms. This is also the type of application for which the methods developed in this thesis are motivated. The thesis investigates and develops both predictive models that can predict properties and simulate these types of systems, and generative models that can propose new potential materials or molecules.

Particular emphasis is put on methods that model data as graphs, and the thesis starts with investigations of graph neural networks (GNNs) designed for predicting material and molecular properties. The performance of these models in the context of high-throughput screenings are put under scrutiny. The performance of GNNs when predicting properties of materials which are only hypothetical and structures which are not completely relaxed is investigated, and the insights are used to suggest a workflow that combines machine learning and conventional high-throughput methods. Additionally, an investigation of so-called knowledge distillation in the context of GNNs for systems of atoms has been performed. This study proposes some simple techniques for improving the performance of this type of GNNs, without sacrificing speed.

The generative modeling techniques developed in the thesis are both more generally applicable and specifically targeting the materials science domain. Among the general methods, the thesis investigates a type of generative autoregressive models where the generation order is a random variable, and develops discriminator guidance for such models. Additionally, a new sequential Monte Carlo algorithm, DDSMC, is developed for general Bayesian inverse problems. A dedicated materials science model, WyckoffDiff, is developed, utilizing a description of materials that explicitly encode information of their symmetries, with the aim of facilitating generation of materials with strict symmetrical properties.

While predictive and generative models can be useful on their own, the study on WyckoffDiff also highlights how they can be used together as parts of a materials discovery pipeline, with predictive models predicting the properties of the materials generated by a generative model like WyckoffDiff.

Discovering new crystalline materials lies at the frontier of modern materials science, driving innovation in energy storage, catalysis, semiconductors, and beyond. The vastness of the chemical and structural space poses a profound challenge: the number of possible atomic arrangements grows prohibitably large with system size and composition. Traditional first-principles methods such as density functional theory (DFT) have revolutionized materials discovery, but their high computational cost limits large-scale exploration. This work addresses the combinatorial bottleneck by bringing together two complementary dimensions of modern materials discovery: data-driven predictions using machine learning and high-performance computing.

The work presented in this thesis builds on a symmetry-aware representation of crystal structures called protostructures, based on Wyckoff positions: a coordinate free description of symmetry related atomic sites. This formulation transforms the continuous space of atomic coordinates into a discrete and combinatorially enumerable one. We developed a machine learning model, Wren, which is trained on this representation to provide fast estimates of stability and guide exploration toward promising regions of structural space. A GPU-accelerated workflow using machine-learning-based interatomic potentials and parallelized screening allows for the evaluation of billions of candidate structures within practical timeframes.

Building on this framework, the presented work enumerates 39 billion binary and ternary compounds spanning the chemical space from lithium to bromine, identifying over 88,000 new structural prototypes, and about half a million new crystal structures within a stability limit of 100 meV/atom. The approach is further applied to experimentally unresolved powder diffraction data, where it reconstructs crystal structures consistent with measured patterns, demonstrating the workflow’s ability to uncover physically realizable materials beyond known prototypes.

To explore even broader regions of structural complexity, this work introduces WyckoffDiff, a diffusion-based generative model that produces novel, symmetry-consistent protostructures beyond the training distribution, some predicted to be thermodynamically stable.

Since pretrained interatomic potentials form the foundation of this work, their quality was examined through two complementary studies. The first benchmarks their accuracy in reproducing mixing enthalpies across disordered alloys. The second investigates how these potentials capture the topology of potential energy surfaces by probing energy variations along symmetry-constrained pathways, showing how different machine-learning potentials represent local minima and saddle points, and other artifacts. These two benchmarks provides insight into their reliability for structure prediction, and the resulting findings informed the selection and parametrization of models used throughout our screening framework.

Altogether, the work presented in this thesis demonstrates that the combination of coarse grained screening, ML-based interatomic potentials, and high-performance computing can dramatically accelerate the discovery of previously unseen crystal structures. The framework presented in the thesis expands the boundaries of computational materials discovery and represents a step toward a large-scale, perhaps even comprehensive, mapping of all stable crystal structures permitted by chemistry and symmetry.

Materials design has evolved from experimental trial-and-error, to model-based theoretical science and physics-based computational methods, and have now entered the paradigm of data-driven methods. During the past decades physics-based methods such as density functional theory (DFT) has been the driving force of computational materials design, and have been used to produce large sets of materials data by academic groups across the world. Since different databases use different technologies and data formats, the data production has resulted in scattered heterogeneous materials database designs. Data driven design of materials relies on data and methods operating on the data such as machine learning, neural networks, and analysis tools. To enable the full potential of the emerging data operational methods, especially the learning-based methods which commonly improves in correlation to the amount of training data, we need to make the data collectively accessible.

Database integration and interoperability are used to make the data collectively and viably accessible, where ontologies are key to, e.g., make data storages machine interpretable and unification of heterogenous data descriptions. This thesis presents an ontology for units of measures specifically targeting challenges of handling units across databases of computational and experimental data. The ontology is created using definition files part of the property definitions in release 1.2, from the community-driven OPTIMADE standard for a common application programming interface (API) for materials databases. The resulting ontology allows addressing data integration challenges encountered in that effort, which were unsolvable with available tools. These challenges include reference to specific and generalized units that have changed over time, using relevant symbols for different scientific domains, specification of relationships between units, and representation of units not part of System of Units (SI) without defining them in SI units or using SI system conventions.

Developments of structure-to-property machine learning and artificial intelligence (AI) mod-els part of data-driven methods has sped-up traditional physics-based simulation methods. Despite the speed increase of structure-to-property predictions, all possible materials is a massive space which is intractable to screen exhaustively, which leads to the idea of in-verse materials design using property-to-structure prediction. Generative AI models are promising candidates for property-to-structure prediction. Most existing crystal structure generative AI models does not account for symmetry, resulting in generation of high concentration of low-symmetry crystal structures. This thesis presents a generative AI model, capable of generating novel and stable symmetry-based descriptions of crystal structures, which inherently accounts for crystal structure symmetry.