Big-Data Science in Porous Materials: Materials Genomics and Machine Learning

Kevin Maik Jablonka; Daniele Ongari; Seyed Mohamad Moosavi; Berend Smit

Source-linked AI summary

Big-Data Science in Porous Materials: Materials Genomics and Machine Learning

Kevin Maik Jablonka, Daniele Ongari, Seyed Mohamad Moosavi, Berend Smit

arXiv:2001.06728v3Jan 18, 2020cond-mat.mtrl-scics.LG

TL;DR

Porous-material research faces an enormous and difficult-to-screen design space. This review surveys big-data and supervised machine-learning methods for representing, modeling, optimizing, and applying data to porous materials, including gas storage and separation, stability, electronic properties, and synthesis. It also identifies hypothetical-database limitations and discusses emerging techniques for the field.

Problem
Porous-material libraries are too large for brute-force experimental or computational screening, motivating methods that can analyze abundant data and discover complex correlations.
Method
The review synthesizes principles of big-data science and supervised machine learning, covering data collection, dataset augmentation and selection, feature representations, model optimization, performance quantification, and applications to porous materials.
Results
The review shows that machine learning has been applied across gas storage and separation, material stability, electronic properties, and synthesis, with applications including gas-adsorption prescreening and uncertainty-guided force-field refinement.
Takeaways & Limitations
Big-data science provides a framework for studying the large and varied space of porous materials, while physical insight can guide model development.
Takeaways & Limitations
Hypothetical databases cover only a limited part of the design space because they use finite sets of linkers and nodes and may not represent the real world.

Abstract

from arXiv · show

By combining metal nodes with organic linkers we can potentially synthesize millions of possible metal organic frameworks (MOFs). At present, we have libraries of over ten thousand synthesized materials and millions of in-silico predicted materials. The fact that we have so many materials opens many exciting avenues to tailor make a material that is optimal for a given application. However, from an experimental and computational point of view we simply have too many materials to screen using brute-force techniques. In this review, we show that having so many materials allows us to use big-data methods as a powerful technique to study these materials and to discover complex correlations. The first part of the review gives an introduction to the principles of big-data science. We emphasize the importance of data collection, methods to augment small data sets, how to select appropriate training sets. An important part of this review are the different approaches that are used to represent these materials in feature space. The review also includes a general overview of the different ML techniques, but as most applications in porous materials use supervised ML our review is focused on the different approaches for supervised ML. In particular, we review the different method to optimize the ML process and how to quantify the performance of the different methods. In the second part, we review how the different approaches of ML have been applied to porous materials. In particular, we discuss applications in the field of gas storage and separation, the stability of these materials, their electronic properties, and their synthesis. The range of topics illustrates the large variety of topics that can be studied with big-data science. Given the increasing interest of the scientific community in ML, we expect this list to rapidly expand in the coming years.

1 Introduction

Porous-material libraries now contain tens of thousands of synthesized structures and many predicted materials, creating both opportunities for materials design and challenges for data-driven analysis. This review introduces machine learning approaches for porous materials and places them within the broader context of materials science and chemistry.

Over 10,000 porous and 80,000 non-porous MOFs have been synthesized, alongside COFs, PPNs, zeolites, and related porous materials.
The growing number of synthesized and predicted structures creates challenges in data management and in discovering new science.
The review examines how machine learning has been applied to porous materials while introducing relevant techniques at each stage of the workflow.
Its focus is materials science and chemistry, particularly porous materials, while broader machine-learning discussions are covered by other reviews.

2 The Machine Learning Landscape

The review presents machine learning as a data-driven addition to scientific discovery, motivated by the enormous design space and data volume of porous materials. It outlines a workflow for applying, evaluating, interpreting, and physically guiding ML models across materials and chemistry.

Motivation: Millions of possible materials and extensive synthesized and predicted databases make brute-force materials screening impractical.The resulting challenge is not only data management but also discovering new science from the available structures and data.
Motivation: Big-data science uses statistically significant patterns in large datasets to support data-driven discovery beyond empirical, theoretical, and simulation-based approaches.The review frames this as the “fourth paradigm of science.”
Machine Learning Workflow: The ML workflow begins by defining the phenomenon and target variable, then generating, exploring, and featurizing reliable data with sufficient design-space coverage.Problem definition determines whether the task is regression or classification and guides data preparation.
Applications and Scope: The review surveys ML applications in porous-material synthesis and applications while emphasizing guidance, caveats, and practical resources for implementing workflow steps.It focuses on supervised learning but also discusses broader ML categories and generative applications.
Theory-Guided Data Science: Physical insight can guide descriptors, model structure, constraints, and hybrid approaches, helping theory-guided data science seek correct answers for scientifically appropriate reasons.One example builds potential-energy-surface models from forces to respect energy conservation.
Machine Learning Workflow: Model evaluation and interpretation should test alternative explanations, including architecture complexity, hyperparameter optimization, and whether learned relationships exceed random-label performance.The review calls this approach strong inference and connects it to ablation and randomization tests.

3 Selecting the Data: Dealing with Little, Imbalanced and Non-Representative Data

Reliable, representative training data is more important than simply increasing model sophistication. The review emphasizes sampling strategies that cover chemical space while recognizing that hypothetical databases and random splits can misrepresent experimental materials.

Data quality: Training data quality is the first priority in ML because unsuitable data can undermine predictions regardless of algorithm choice.The review frames this as “garbage in, garbage out” and states that data matters more than algorithms.
Sampling and splitting: Random data splitting can miss sparsely populated chemical-space regions and distort class distributions, especially when datasets are small or imbalanced.Visualization, stratification, and deliberate sampling help expose or mitigate these problems.
Non-representative data: Hypothetical COF, MOF, and zeolite databases are large and self-consistent but cover only a limited, potentially non-real-world portion of design space.Their finite sets of linkers and nodes constrain coverage and representativeness.
Farthest point sampling: Farthest point sampling greedily selects diverse structures by maximizing their minimum distance from already selected points, avoiding redundancy more efficiently than exact maximum-diversity optimization.The exact maximum diversity problem has factorial computational cost, so greedy FPS is used in practice.
Design of experiments: Design-of-experiments methods, including max-min Latin hypercube sampling, seek efficient coverage of relevant combinations and reaction conditions.These methods are used to construct informative initial training sets with limited experiments or calculations.
Active learning: Active learning lets the model choose data that are most effective for improvement, which is valuable when experiments or high-accuracy calculations are expensive.This avoids relying on brute-force generation of large reference libraries.

4 What To Learn From: Translating Structures Into Feature Vectors

Materials must be translated into feature vectors before classical ML can learn structure–property relationships. Descriptor choices range from engineered fingerprints and chemically meaningful features to neural-network representations learned directly from raw structural data.

Descriptor design: Feature columns encode structures as descriptors, and selecting more descriptors is not always better because different tasks require different structural information.Local fingerprints may suit potential-energy surfaces, whereas pore shape and size can better support structure–property interpretation.
Representation learning: Neural networks can learn representations from unstructured inputs such as atomic positions and charges, reducing reliance on highly engineered descriptors.This advantage requires large amounts of data and may capture regularities unavailable to classical descriptors.
MOF identifiers: MOFid decomposes MOFs into building blocks and encodes linkers with SMILES to create unique identifiers that simplify data mining.The scheme addresses naming and non-uniqueness problems encountered in text-mining workflows.
Descriptor limitations: MOF fingerprinting can require bond and bond-order assignments, which are difficult for disordered or incorrectly protonated experimental structures.This limits the reliability of such fingerprints for some solid-state datasets.

4.1 Descriptors

Useful descriptors should preserve relevant physical symmetries, avoid degeneracy, generalize across compositions, and remain computationally efficient. Descriptor locality is powerful for some properties but can fail when pore-scale or long-range physics dominates.

Descriptor requirements: An ideal descriptor is unique, transformation-invariant, cross-element generalizable, efficient to calculate, and continuous when differentiability is needed.These requirements aim to prevent irreducible descriptor ambiguity, limit system-size restrictions, and preserve efficient ML workflows.
Symmetry: Crystal descriptors should respect periodicity, translational, rotational, and permutation symmetries so equivalent structures receive consistent representations.Encoding symmetries directly is described as more robust and efficient than relying only on data augmentation.
Symmetry: Hard-coding descriptor invariances removes alignment problems when comparing systems.The review links this benefit to descriptor-level treatment of structural transformations.
Dimensionality: The curse of dimensionality makes decision boundaries difficult to learn because high-dimensional datasets leave much of the space uncovered.In 100 dimensions, nearly the full edge length is needed to capture 10% of a unit hypercube’s volume.
Locality: Chemical locality approximates a material property as a sum of contributions from local atom-centered environments, an approach commonly used for potential-energy surfaces.The approximation is motivated by the nearsightedness principle of electronic matter and can support size-extensive modeling.
Locality: For gas adsorption, pore shape, size, and accessibility may require global pore-centered descriptors rather than atom-centered local environments.Persistent-homology fingerprints and pore metrics are cited as suitable alternatives.
Locality: Locality-based descriptors can miss long-range effects, whose correct treatment remains a major research challenge.The review notes that typical local environments use cutoffs around 6 Å.

4.2 An Overview of the Descriptor Landscape

The descriptor landscape spans local atom-centered and global structure-wide representations, with many approaches built from related many-body correlation ideas. Applications include symmetry functions, SOAP, Voronoi-based environments, and persistent-homology fingerprints for porous-material prediction.

Local and global descriptors: Descriptors divide broadly into local representations of chemical environments and global representations of complete structures.The review presents this distinction as the main organizing principle for the descriptor landscape.
Correlation-based descriptors: Many fingerprinting methods are variations of many-body correlation functions projected onto different mathematical bases and correlation orders.Higher-order descriptors can incorporate triples or additional terms, but may converge differently as complexity increases.
Local descriptors: Local descriptors commonly use smooth cutoff functions, while Voronoi tessellations can assign nearest neighbors and coordination numbers.Cutoffs define the included environment; Voronoi cells define regions closer to a central atom than to others.
Symmetry functions: Behler–Parrinello symmetry functions encode two-body distance information and three-body angular information within a cutoff environment.Their hyperparameters include Gaussian shifts, widths, and the cutoff radius, which should be optimized or converged.
Symmetry functions: Symmetry-function fingerprints are translationally and rotationally invariant, while neighbor summation provides permutation invariance.These properties arise from dependence on internal coordinates, internal angles, and order-independent summation.
SOAP: SOAP represents local environments with smooth atomic densities and directly supplies a similarity measure for kernel-based learning.Its extension to tensorial properties enabled molecular polarizability models.
Topological descriptors: Persistent-homology descriptors record the creation and destruction of topological features during a radius-based filtration, and barcode similarity can rank or predict porous-material performance.The approach was used with random forests and with kernels for kernel ridge regression in methane-storage applications.

4.3 Feature Learning

Feature learning transforms raw descriptors and selects compact, informative feature sets because brute-force descriptor combinations are computationally intractable. The review surveys filter, wrapper, sparse-optimization, clustering, and stability-based approaches, including applications that discover interpretable materials descriptors.

Feature engineering: Feature engineering forms new descriptors by combining or mathematically transforming raw features, allowing domain knowledge to enter the modeling process.Examples include logarithmic transformations and physically meaningful ratios.
Feature selection: Because exhaustive descriptor selection is NP-hard, practical methods use filters, wrapper iterations, or approximations and heuristics.Simple filters can rank features by correlation, mutual information, or single-feature fitting error, but they miss interaction effects.
Wrapper approaches: Recursive feature elimination repeatedly removes the least important features, while weight pruning removes features with low fitted weights; both can be computationally expensive.Feature importance may change after each elimination in recursive procedures.
Wrapper approaches: Reducing the feature set from ca. 150 to 22 features with RF-RFA reduced test-set MAE from 9.5 kcal/mol to 5.5 kcal/mol.The reported result links the selected feature-set size and the corresponding test error.
Sparse methods: LASSO uses the l1 norm as a convex approximation to the NP-hard l0 objective, penalizing high weights to promote sparsity and avoid overfitting.Compressed sensing gives conditions under which the l0 and LASSO solutions likely coincide and is suited to sparse, noisy materials problems.
Sparse methods: Compressed sensing enabled construction of massive 10^9-descriptor sets and discovery of new perovskite tolerance factors and interpretable quantum spin-Hall-insulator descriptors.The descriptors were generated by applying algebraic functions to primary features such as atomic radii.
Stability and compression: Clustering correlated features before projected-gradient selection and bootstrapping can compress the feature set and make feature selection more stable.Stability selection retains features selected across bootstrap samples to counter chance correlation.
Feature transformation: Feature distributions and scales may require transformation because units or representations can hinder machine learning and interpretability.The review highlights logarithmic and reciprocal scales as chemically meaningful alternatives to raw values.

5 How To Learn: Choosing a Learning Algorithm

Choosing a learning algorithm requires matching model families to data and understanding how flexibility affects bias, variance, and irreducible error. The review presents the algorithm landscape and emphasizes the classical bias-variance framework alongside emerging deep-learning behavior.

Algorithm selection: The review gives a non-exhaustive overview of the supervised learning algorithm landscape and generally does not distinguish classification from regression.Many algorithms can be formulated for both problem types.
Bias-variance trade-off: The bias-variance decomposition expresses total error as squared bias, variance, and irreducible Bayes error.The decomposition can be derived by rewriting the mean-square-error cost function.
Bias-variance trade-off: Variance reflects finite-training-size fluctuations, whereas bias is the error remaining with infinitely many training points and reflects model expressivity limits.Irreducible error cannot be removed by choosing a better model and can arise from noise in training data.
Bias-variance trade-off: Highly flexible models tend toward high variance, while simple regularized linear models tend toward high bias; an intermediate flexibility can minimize test error.The review describes this intermediate region as a test-error “sweetspot.”
Deep learning: Deep-learning research examines why highly overparameterized models can still achieve good test error, with double descent proposed after the jamming transition.The proposed explanation involves access to a larger function space containing simpler interpolating functions.
Deep learning: Neural networks are presented as especially suited to large, unstructured data or minimally preprocessed feature sets because they can learn representations.Message passing is identified as a representation-learning approach in chemical science.

5.1 Lots of (Unstructured) Data (Tall Data)

Neural networks are useful for large, unstructured materials datasets because they scale well and can learn representations directly from structured inputs. The review describes feed-forward, high-dimensional, message-passing, recurrent, and convolutional architectures and their materials applications.

Neural-network architectures: Deep neural networks are promising for large materials datasets that lack highly engineered features, including images, spectra, coordinates, and atom identities.Their representation-learning capability reduces reliance on manually designed descriptors.
Feed-forward networks: Feed-forward networks approximate functions through chained matrix evaluations, with learned weights, biases, and nonlinear activation functions across layers.Multiple nonlinear layers constitute deep learning.
Neural-network architectures: Neural networks scale favorably because stochastic gradient descent avoids expensive O(n^3) matrix inversion and can train using minibatches.Their expressivity and benign scaling support massive unstructured datasets.
Potential-energy surfaces: High-dimensional neural-network potentials encode each atom’s local environment as a fingerprint and can model potential energy surfaces efficiently, accurately, and reactively.The approach addresses scaling and fundamental-invariance limitations of earlier neural-network potentials.
Message-passing networks: Message-passing networks construct problem-relevant descriptors by recursively incorporating interactions from increasingly complex neighboring environments.Each environment initially represents one atom and is refined through successive interactions.
Convolutional networks: Convolutional neural networks slide shared filters over images or patterns, use pooling to compress representations, and can extract higher-level features.Pooling provides approximate translational invariance by reducing sensitivity to small translations.
Convolutional networks: CNNs have been used to classify crystal structures from two-dimensional diffraction patterns and to predict methane deliverable capacity in COFs from XRD patterns.The review also cites applications to steel microstructures and periodic-table representations.

5.2 Limited Amount of (Structured) Data (Wide Data)

With limited structured data, simple and kernel-based models can provide strong baselines, while kernel methods extend linear models to nonlinear relationships. Bayesian formulations additionally provide uncertainty estimates useful for deciding when new data are needed.

Kernel-based models can equal or outperform neural networks on structured data, especially when training data are limited.The review recommends implementing the simplest model first to establish a baseline and verify the modeling infrastructure.
80 sampled points left KRR unable to model the Lennard-Jones potential’s strong repulsion because that region lacked training examples.The related GPR experiment likewise reports failure when training data omit important regions.
KRR uses kernel functions to avoid explicit high-dimensional feature expansion, with computation scaling by data points rather than features.Kernel methods implicitly map data into higher-dimensional spaces through inner products.
Regularization stabilizes kernel solutions when direct matrix inversion is ill-posed, unstable, or non-unique.Adding λI produces the regularized coefficient solution instead of relying on K^-1 alone.
Kernel methods can combine different-fidelity data to predict high-fidelity properties and support multitask prediction across regression tasks.Examples include using GGA data to predict HSE06 band gaps and predicting many properties from one kernel.
Bayesian learning produces posterior distributions and natural prediction-uncertainty estimates that support active learning.In one application, uncertainty triggered DFT calculations and force-field refinement when predictions were too uncertain.

6 How To Learn Well: Regularization, Hyperparameter Tuning, and Tricks

Effective learning with limited or irregular materials data requires careful hyperparameter optimization, regularization, and training controls. The review presents methods that improve efficiency, generalization, and robustness across model classes.

6.1 Hyperparameter Tuning: Hyperparameter tuning is expensive because each candidate requires model training and validation, while grid search wastes prior evaluation information.The number of grid evaluations grows exponentially with the number of hyperparameters.
6.1 Hyperparameter Tuning: Bayesian hyperparameter optimization is more efficient than grid search because sequential methods use previous evaluations to select promising new configurations.SMBO balances exploration of uncertain regions with exploitation of configurations predicted to perform well.
6.2 Regularization: Chemical and materials datasets are often ill-posed, noisy, sparse, or wide, making regularization important for controlling model complexity.Physical or chemical insights can also provide especially powerful regularization constraints.
6.2.1 Explicit Regularization: Adding an Term or Layer: Lasso regularization promotes sparse solutions, whereas ridge regularization smoothly shrinks weights and encourages smoother solutions.Both recover the unregularized solution as λ →0 and force it toward zero as λ →∞.
6.2 Regularization: Dropout, pruning, and early stopping reduce overfitting in neural networks and tree models by limiting effective model complexity.Early stopping monitors training and validation errors and halts training when validation error no longer decreases.
6.2 Regularization: Stochasticity underlies several regularization techniques, including noise injection, randomly dropping layers, batch normalization, and minibatch SGD.These methods make training or predictions depend on random perturbations or batch composition, similarly to averaging varied models.

7 How to Measure Performance and Compare Models

Model evaluation and selection should reflect the data regime, task objective, class balance, extrapolation needs, and computational resources. The review emphasizes leakage-free validation, appropriate metrics, representative sampling, uncertainty estimation, and ablation studies.

Evaluation objectives: Performance estimates should target unseen-data behavior, rather than relying on a single global metric such as MAE.Global metrics depend on data-point distributions and may not identify top-performing materials or account for training time and data requirements.
Evaluation objectives: Model evaluation methodology depends on dataset size, problem setting, extrapolation requirements, and available computational resources.The review contrasts k-fold cross-validation for little data with three-way holdout evaluation for big data.
Holdout splits and cross-validation: Random splitting can reduce training data, create pessimistic bias, and alter class proportions between training and test sets, especially for small datasets.Stratification preserves class proportions across sets and can be applied to regression by binning the target range.
Holdout splits and cross-validation: K-fold cross-validation gives each example a chance to participate in training, while leave-one-out cross-validation minimizes wasted data but has high computational burden and variance.The review reports k = 10 as an empirical compromise between bias and variance.
Holdout splits and cross-validation: Cross-validation must precede filtering, feature selection, and standardization to prevent test-data leakage and biased feature choices.Hyperparameter tuning should use validation data rather than the test set, which remains the surrogate for unseen-data performance.
Model comparison: Ablation studies isolate whether improvements arise from architecture, optimization, or hyperparameter choices rather than from an apparently more complex model.The review notes that better hyperparameter optimization has sometimes explained improvements attributed to model architecture.

8 How to Interpret the Results: Avoiding the Clever Hans

Interpretability methods help inspect model reasoning, identify structure–property relationships, and detect dataset or modeling problems. The review compares intrinsic and post-hoc approaches while emphasizing their assumptions, interaction effects, and vulnerability to biased data.

Motivation: Interpretability can expose model reasoning, debug dataset problems, and support extraction of design rules, especially with high-level physical features.The review frames this as a response to black-box predictions that otherwise offer little scientific understanding.
Caveats: Interpretations remain crude approximations of nature unless the underlying model has demonstrated predictive ability.The review also notes that “interpretable” has no single well-defined meaning.
Intrinsic interpretability: Simple models provide intrinsic interpretability but may lack expressive power when the problem is complex or the feature set is inadequate.Generalized additive models retain feature-wise analysis while representing nonlinear relationships through smooth functions.
Intrinsic interpretability: GAMs require explicit interaction features unless a modified formulation includes pairwise interactions by default.Their additive structure supports feature-wise analysis but does not automatically represent arbitrary interactions.
Post-hoc techniques: Partial dependence plots visualize average feature effects and interactions but assume plotted features are independent of the others, which is their biggest limitation.Permutation importance can face related problems with correlated features, although correlated features may be permuted together.
Auditing models: Randomly selected reaction conditions outperformed human-selected literature conditions for predicting crystal formation because they sampled feature space better.The literature conditions were anthropogenically biased, with chemistry-of-amine features forming an exception to otherwise linear correlations.

9 Applications of Supervised Machine Learning

The review surveys supervised machine-learning applications across porous-materials gas storage and separation, stability, electronic properties, and synthesis. These studies use descriptors and models to accelerate screening, predict material and process properties, identify uncertainty or data errors, and guide synthesis.

Scope and limitations: Hypothetical-material databases cover only a limited, potentially unrepresentative portion of design space, so observed trends may reflect dataset limitations rather than underlying physics.This limitation is especially relevant when models trained on large hypothetical structures are applied to experimental materials.
Gas storage and separation: Increasing heat of adsorption while decreasing porosity increased CO2/N2 selectivity in a model linking separation performance to material properties.
Gas storage and separation: 945 of the top 1,000 MOFs were correctly identified while only 10% were flagged for GCMC investigation, enabling prescreening of low-performing materials.The classifier used atomic-property-weighted RDF descriptors for carbon dioxide adsorption.
Gas storage and separation: Models predicted whether materials reached 95% CO2 purity and 90% CO2 recovery, but not the parasitic energy required for sorbent regeneration and CO2 compression.Nitrogen adsorption properties were most important for predicting the 95/90-PRT target.
Electronic properties: Ensemble classifiers assigned MOF oxidation states with high predictive performance, quantified uncertainty, and detected errors in the underlying training data.Features included descriptions of local coordination-environment geometry.

10 Outlook and Concluding Remarks

The outlook emphasizes comprehensive ML practice, reproducibility, data sharing, and emerging approaches for porous-materials research. It highlights transfer learning, multifidelity optimization, synthesis data, and provenance-rich workflows as foundations for future progress.

10 Outlook and Concluding Remarks: The review provides a comprehensive overview of state-of-the-art ML in materials science while highlighting material-specific caveats and promising underused techniques.It also discusses methods that may address issues identified throughout the review.
10.1 Automatizing the Machine Learning Workflow: Automated tools such as chemml and automatminer can lower the computational entry barrier by handling descriptors, feature selection, training, and cross-validation.These tools also help ML practitioners expedite tedious tasks.
10.2 Reproducibility in Machine Learning: Reproducibility remains essential because unpublished settings, inconsistent metrics, and missing workflow details hinder replication and comparability.The authors recommend rigorously reporting all input and output files, workflows, hyperparameters, and metrics.
10.2 Reproducibility in Machine Learning: Materials Cloud and AiiDA can expose complete provenance, source workflows, computed data, and parameters for reproducible COF screening and later dataset combination.The workflow covers structure optimization, charge computation, GCMC simulations, adsorption properties, and carbon-capture ranking.
10.3 Transfer Learning and Multifidelity Optimization: Transfer learning and multifidelity optimization address settings where accurate experimental or high-level computational labels are scarce or expensive.Transfer learning was reported to generally improve predictive performance when DFT labels support prediction of experimental properties.
10.5 The Future of Big-Data Science in Porous Materials: Sharing failed and partially successful MOF synthesis experiments could enable ML to relate linker and metal-node choices to crystallization conditions.Such data could support predictions of optimal synthesis conditions for novel MOFs.

Author Information

The authors are researchers in chemistry, chemical engineering, physics, and data science whose work centers on computational and data-driven porous-materials research. Their applications include molecular adsorption, catalysis, and energy-related material design.

Author Information: Kevin Maik Jablonka investigates data-driven methods for designing and discovering materials for energy-related applications.He studied chemistry and data science and is a PhD student at EPFL.
Author Information: Daniele Ongari studies microporous MOFs and COFs computationally to assess their performance for molecular adsorption and catalysis.He completed his PhD under Berend Smit at EPFL in 2019.
Author Information: Seyed Mohamad Moosavi focuses on computational and data-driven design and engineering of novel materials for energy-related applications.His background includes mechanical engineering and chemistry and chemical engineering.
Author Information: Berend Smit’s research centers on molecular simulation techniques, especially for energy-related applications.He is Professor of Computational Chemistry at the University of Amsterdam.