François-Xavier Coudert

There is mounting evidence that different amorphization routes can produce metalorganic framework (MOF) glasses with distinct structures and properties, yet the microscopic origins of these differences remain difficult to establish experimentally. Focusing on the two main synthesis methods, which are melt-quenching and ball-milling, we use molecular dynamics simulations with a machine-learning interatomic potential to compare different glasses obtained from ZIF-4. While both routes largely preserve tetrahedral zinc coordination, mechanical deformation generally produces more distorted local environments, showing that average coordination numbers alone do not capture the structural differences between these glasses. At smaller applied strain, mechanical glasses retain substantial memory of the crystalline network topology, despite pronounced changes in ring geometry. Larger strains erase this topological memory, yielding ring-size distributions similar to those of melt-quenched glasses, while differences in local geometry and porosity persist. Simulations of larger systems reveal a heterogeneous amorphization mechanism, involving the formation and growth of disordered domains. These findings from in silico amorphization support process-dependent polyamorphism in ZIF glasses and identify local geometry, topological memory, and spatial heterogeneity as complementary features needed to characterize their structure and understand the influence of mechanical processing.

Machine-learned interatomic potentials (MLIPs) have emerged as a strategy to accelerate molecular simulations, offering the promise of quantum chemical accuracy at a cost close to that of classical force fields. They are commonly trained on reference data obtained at the DFT level, which is itself a sufficiently affordable method to generate thousands of configurations and associated energies and atomic forces. However, this approach suffers from a major underlying challenge, which is rarely addressed: the choice of exchange–correlation functional. Although it is understood to have a crucial impact on the accuracy of the description of interactions and therefore on the results of the molecular simulations, it has not been systematically evaluated in the past. To go beyond simple benchmarks on selected configurations, we set out to understand the influence of exchange–correlation (XC) functionals on the training of MLIPs and, more importantly, on the physical properties of the resulting condensed matter systems. As a complex chemical system to test this, we selected borosilicate glasses of a wide variety of compositions. We produced reference data sets of borosilicate glasses with ten different chemical compositions, with six different XC functionals at the GGA, meta-GGA, and hybrid levels. We then trained MLIPs for all of them, produced glasses by melt-quenching, and compared their physical and structural properties with available experimental data. We show that there are significant differences between the different functionals. While the influence on bulk properties such as glass densities is minor, fine structural properties are more sensitive to the choice of functional, in particular when it comes to the coordination of boron atoms. Our results show how MLIPs can help us better understand the strengths and weaknesses of DFT functionals, where investigation through direct ab initio molecular dynamics was too expensive in the past. This also highlights that MLIPs depend on the ground truth on which they are trained, a fact often overlooked in practice, and this influence may be drastic for specific physical properties.

Machine-learnt interatomic potential (MLIP) has emerged as a strategy to accelerate molecular simulations, offering the promise of quantum chemical accuracy at a cost close to that of classical force fields. They are commonly trained on reference data obtained at the DFT level, which is itself a sufficiently affordable method to generate thousands of configurations and associated energies and atomic forces. However, this approach suffers from an elephant in the room, which is rarely addressed: the choice of exchange–correlation functional. Although it is understood to have a crucial impact on the accuracy of the description of interactions and therefore on the results of the molecular simulations, it has not been systematically evaluated in the past. To go beyond simple benchmarks on selected configurations, we set out to understand the influence of exchange–correlation (XC) functional on the training of MLIPs, and more importantly on the physical properties of the resulting condensed matter systems. As a complex chemical system to test this, we selected borosilicate glasses of a wide variety of compositions. We produced reference data sets of borosilicate glasses with nine different chemical compositions, with five different XC functionals at the GGA, meta-GGA and hybrid levels. We then trained MLIPs for all of them, produced glasses by melt-quenching, and compared their physical and structural properties with available experimental data. We show that there are significant differences between the different functionals. While the influence on bulk properties such as glass densities is minor, fine structural properties are more sensitive to the choice of functional, in particular when it comes to the coordination of boron atoms. Our results show how MLIPs can help us better understand the strengths and weaknesses of DFT functionals, where investigation through direct ab initio molecular dynamics was too expensive in the past. It also highlights that MLIPs depend on the ground truth on which they are trained, a fact often overlooked in practice, and this influence may be drastic on specific physical properties.

The growing number of hypothetical metal–organic frameworks (MOFs) has created a need for reliable methods to identify candidates that can be synthesized experimentally. This Viewpoint examines current computational approaches to the evaluation of MOF synthesizability, grouping them into thermodynamic, strain- and mechanics-based, and data-driven methods. Thermodynamic studies evaluate formation enthalpies or free energies, stability relative to competing phases, and can include solvent effects. Structural approaches quantify linker deformation and mechanical stability as a proxy for viability of a hypothetical structure. Data-driven machine learning models use geometric, topological, or chemical relationships to known materials as a way to estimate synthesizability. We review the recent advances of these approaches, their strengths and their limitations, noting that at the present time, there is no silver bullet to solve this open challenge. Future progress will require combining complementary methods and using consistent experimental data to refine computational models and allow the identification of candidates for synthesis in screening studies.

Zeolitic imidazolate frameworks (ZIFs) are a family of metal–organic frameworks that feature metal centers tetrahedrally linked to imidazole-based ligands and adopt zeolite-like topologies. ZIFs formed by zinc cations and imidazolate linkers exhibit a remarkable degree of polymorphism, which can be modulated by varying synthesis parameters or thermodynamic conditions (i.e., temperature and pressure). Computer simulations provide a unique way of studying these materials and their phase transitions from the microscopic standpoint, revealing their underlying molecular mechanisms. However, studying these mechanisms requires to be able to classify the phase of each molecular entity in an agnostic and automatic fashion, which is particularly challenging when the two phases involved are structurally very similar. In this work, we systematically study neural network classifiers to classify ZIF phases on-the-fly during molecular dynamics simulations. We test a variety of input features, differing both in the dimensionality and nature of the descriptors and in the kind of force field used for building the training/testing database. We reveal that even with low-dimensional descriptors, the classification is highly accurate, while the use of high-dimensional descriptors leads to an even better performance. Training the classifier with configurations coming from different force fields, we can remove force field bias and enhance the classifier performance and general applicability. Finally, we apply our classifiers to reveal mechanistic details of the ZIF-4-cp → ZIF-4-cp-II phase transition.

Effective industrial separation processes based on nanoporous materials rely heavily on the kinetic properties of the guest molecules in the porosity, yet computational screening studies in the past have focused mostly on the adsorption (or co-adsorption) thermodynamics. Moreover, traditional computational chemistry methods for predicting diffusion coefficients, such as classical molecular dynamics simulations, are often too computationally demanding to be used in high-throughput screening workflows. This work introduces a deep learning framework that overcomes this bottleneck by leveraging 3D voxelized grids of the guest molecule’s Potential Energy Surface (PES) as direct inputs for a 3D Convolutional Neural Network (CNN). Trained on a large dataset of about 7,000 diffusion coefficients for xenon and krypton in MOFs, the deep-learning predictor shows good accuracy with low computational effort. Unlike conventional machine learning approaches that rely on extracted scalar porosity and energy descriptors of the nanoporous materials, our method captures the spatial variations of the hostguest potential energy landscape, enabling a fast estimation of diffusion coefficients. By demonstrating the predictive power of 3D CNNs for guest diffusion prediction on the CoRE MOF 2024 database, we provide a scalable, physically-informed methodology for the multi-scale screening of materials (experimental and hypothetical) and establish a new open-source benchmark dataset to spur development of the field of transport property prediction.

Machine-learnt interatomic potential (MLIP) has emerged as a strategy to accelerate molecular simulations, offering the promise of quantum chemical accuracy at a cost close to that of classical force fields. They are commonly trained on reference data obtained at the DFT level, which is itself a sufficiently affordable method to generate thousands of configurations and associated energies and atomic forces. However, this approach suffers from an elephant in the room, which is rarely addressed: the choice of exchange–correlation functional. Although it is understood to have a crucial impact on the accuracy of the description of interactions and therefore on the results of the molecular simulations, it has not been systematically evaluated in the past. To go beyond simple benchmarks on selected configurations, we set out to understand the influence of exchange–correlation (XC) functional on the training of MLIPs, and more importantly on the physical properties of the resulting condensed matter systems. As a complex chemical system to test this, we selected borosilicate glasses of a wide variety of compositions. We produced reference data sets of borosilicate glasses with nine different chemical compositions, with five different XC functionals at the GGA, meta-GGA and hybrid levels. We then trained MLIPs for all of them, produced glasses by melt-quenching, and compared their physical and structural properties with available experimental data. We show that there are significant differences between the different functionals. While the influence on bulk properties such as glass densities is minor, fine structural properties are more sensitive to the choice of functional, in particular when it comes to the coordination of boron atoms. Our results show how MLIPs can help us better understand the strengths and weaknesses of DFT functionals, where investigation through direct ab initio molecular dynamics was too expensive in the past. It also highlights that MLIPs depend on the ground truth on which they are trained, a fact often overlooked in practice, and this influence may be drastic on specific physical properties.

Because topology plays a key role in many chemical and physical properties of materials, identification of topology from crystalline structures is a common and important task in materials science. We present here a new web application, CrystalNets, whose user-friendly interface allows scientists to identify and visualize the topology of crystals from their atomic structure in CIF format. The software has a lot of options to customize its features, such as detection of bonding, choice of clustering model, type of materials (inorganic, hybrid, etc). It has a default mode with powerful heuristics, and allows the user to easily visualize and check the topology detected against the full structure. We also improved the underlying CrystalNets.jl Julia library, including a systematic algorithm to handle the complex case of unstable nets.

Metal–organic frameworks (MOFs) feature a rich structural diversity, including crystalline, amorphous, and liquid phases of varying topologies. Their structural characterization is often performed either at the local scale (through pair distribution functions, bond angle distributions, etc.) or, for crystalline phases, through topology analysis of the periodic framework—leaving out disordered and amorphous phases. In this work, we develop a computational methodology for the structural characterization of middle-range order in MOFs that is applicable to both crystalline and amorphous phases. We base our method on the statistical analysis of the geometry of the supramolecular framework at the microscopic level, and its evolution during molecular simulation. We analyze the statistics of metal–organic rings, their distribution in size, as well as their geometrical characteristics through mathematical tools derived from polymer physics: radius of gyration, asphericity, and writhe. We show that this advanced characterization can be leveraged for the identification of phases and the detection and analysis of phase transitions.