Chapter 5: Machine Learning and Databases

Learning Objectives

Understand the landscape of MOF databases
Learn how machine learning accelerates MOF discovery
Explore generative AI for designing new MOFs
Appreciate the integration of computation and experiment
Recognize future directions in computational MOF science

5.1 The Need for Computational Approaches

The MOF Discovery Challenge

With over 100,000 MOF structures reported and millions more theoretically possible, experimental screening of all candidates is impossible. Computational methods and machine learning offer a solution.

flowchart TD A[~100,000 Known MOFs] --> B[Millions of
Hypothetical MOFs] B --> C{How to Find
Best MOF?} C --> D[Experimental
Screening] C --> E[Computational
Screening] D --> F[Slow: ~10 MOFs/year
per researcher] E --> G[Fast: ~1 million MOFs
screened in days] G --> H[Top Candidates] H --> I[Targeted
Experiments] style E fill:#27ae60,stroke:#1e8449,color:#fff style G fill:#3498db,stroke:#2980b9,color:#fff

Computational Workflow

Database mining: Search existing MOF databases
Structure generation: Create hypothetical MOF structures
Property prediction: Calculate or predict target properties
Screening: Rank candidates by performance
Experimental validation: Synthesize and test top candidates

5.2 Major MOF Databases

Cambridge Structural Database (CSD) MOF Subset

CSD MOF Collection

Source: Cambridge Crystallographic Data Centre (CCDC)
Content: ~100,000+ MOF structures (as of 2024)
Data: Crystal structures from published literature
Access: Subscription required (academic licenses available)
Features: High-quality, experimentally verified structures

CoRE MOF Database

Computation-Ready, Experimental MOF Database

Source: Northwestern University (Snurr group)
Content: ~14,000 curated structures (2019 version)
Data: Cleaned structures ready for molecular simulation
Access: Free (open access)
Features: Removed solvent, assigned charges, standardized format
URL: github.com/gregchung/gregchung.github.io

QMOF Database

Quantum MOF Database

Source: Georgia Tech (Sholl group)
Content: ~20,000 MOFs with DFT-computed properties
Data: Band gaps, partial charges, electronic properties
Access: Free (open access)
Features: Consistent DFT calculations across all structures

hMOF (Hypothetical MOF) Databases

Computationally Generated MOFs

Content: 100,000+ to millions of hypothetical structures
Generation: Combinatorial assembly of building blocks
Use: High-throughput screening for target applications
Examples: hMOF database, ToBaCCo-generated MOFs

Comparison of Databases

Database	Size	Type	Access	Best For
CSD MOF	~100,000	Experimental	Subscription	Literature structures
CoRE MOF	~14,000	Curated experimental	Free	Molecular simulation
QMOF	~20,000	DFT-computed	Free	Electronic properties
hMOF	100,000+	Hypothetical	Free	Discovery screening

5.3 Machine Learning for MOF Property Prediction

Why Machine Learning?

First-principles calculations (DFT, molecular dynamics) are accurate but slow. Machine learning models trained on computed or experimental data can predict properties in milliseconds.

flowchart LR subgraph Traditional["Traditional Approach"] A1[MOF Structure] --> B1[DFT Calculation] B1 --> C1[Property] B1 -.->|Hours to Days| C1 end subgraph ML["ML Approach"] A2[MOF Structure] --> B2[Feature Extraction] B2 --> C2[ML Model] C2 --> D2[Property] C2 -.->|Milliseconds| D2 end style B1 fill:#e74c3c,stroke:#c0392b,color:#fff style C2 fill:#27ae60,stroke:#1e8449,color:#fff

Feature Engineering for MOFs

ML models require numerical representations (features) of MOF structures:

Structural Features

Geometric: Surface area, pore volume, pore size distribution
Topological: Network topology, node/linker connectivity
Chemical: Metal type, linker functional groups, elemental composition

Advanced Representations

Revised Autocorrelation (RAC): Encodes local chemical environment
Persistent homology: Captures topological features at multiple scales
Graph neural networks: Learn features directly from atomic connectivity
Crystal graph: Represents periodic structure as a graph

Common ML Models for MOFs

Model Type	Strengths	Example Applications
Random Forest	Interpretable, handles mixed features	Gas uptake prediction
Gradient Boosting	High accuracy, feature importance	CO₂/N₂ selectivity
Neural Networks	Complex patterns, large datasets	Multi-property prediction
Graph Neural Networks	Direct structure input, transferable	Stability prediction
Transformer Models	Attention mechanism, state-of-the-art	Property prediction, generation

Example: Predicting Gas Adsorption

ML Workflow for CO₂ Uptake Prediction

Data collection: Gather MOF structures + experimental/computed CO₂ uptake
Feature calculation: Compute surface area, pore volume, chemical descriptors
Model training: Train random forest or neural network on 80% of data
Validation: Test on remaining 20%, evaluate R² and MAE
Screening: Apply model to hypothetical MOF database
Selection: Identify top candidates for synthesis

Success Stories

ML-Guided Discovery of Record CH₄ Storage MOF

Researchers used ML to screen millions of hypothetical MOFs for methane storage. The top predicted candidate was synthesized and achieved record-breaking methane uptake, validating the computational approach.

Stability Prediction

ML models can now predict whether a hypothetical MOF will be thermally stable or water-stable with >85% accuracy, allowing researchers to focus on synthesizable targets.

5.4 Generative AI for MOF Design

Beyond Prediction: Generating New MOFs

While predictive ML screens existing candidates, generative AI can design entirely new MOF structures optimized for specific properties.

flowchart LR A[Target Property
e.g., High CO₂ uptake] --> B[Generative
AI Model] B --> C[New MOF
Structures] C --> D[Validation
Simulation] D --> E[Synthesis &
Testing] style B fill:#9b59b6,stroke:#8e44ad,color:#fff style C fill:#27ae60,stroke:#1e8449,color:#fff

Generative Approaches

Variational Autoencoders (VAE)

VAEs learn a continuous latent space of MOF structures. New MOFs can be generated by sampling from this space or by optimizing in latent space for desired properties.

Generative Adversarial Networks (GAN)

GANs use a generator-discriminator framework to create realistic MOF structures that could plausibly exist.

Diffusion Models

Recent advances apply diffusion models (like those behind image generators) to generate 3D crystal structures, including MOFs.

Language Models for Chemistry

Large language models trained on chemical data can generate SMILES strings for new linkers or suggest synthesis conditions.

MOF Generation Pipelines

Inverse Design Workflow

Define target: Specify desired properties (e.g., high H₂ uptake at 298 K)
Generate candidates: Use generative model to propose structures
Filter: Remove chemically unreasonable structures
Predict properties: Use ML to estimate performance
Simulate: Run molecular dynamics on top candidates
Synthesize: Attempt synthesis of best structures

Challenges in Generative MOF Design

Synthesizability: Generated structures may not be synthetically accessible
Stability: Ensuring generated MOFs are thermodynamically stable
Validity: Maintaining chemical validity (correct bonding, stoichiometry)
Diversity: Avoiding mode collapse to similar structures

5.5 Molecular Simulation Methods

Grand Canonical Monte Carlo (GCMC)

GCMC simulations predict gas adsorption isotherms by simulating gas molecules inserting, deleting, and moving within MOF pores at fixed temperature and chemical potential.

What GCMC Predicts

Gas uptake at various pressures and temperatures
Adsorption isotherms
Selectivity between different gases
Adsorption sites and mechanisms

Molecular Dynamics (MD)

MD simulations track the motion of atoms over time, providing insights into:

Diffusion of guest molecules through pores
Framework flexibility and dynamics
Thermal stability
Transport properties

Density Functional Theory (DFT)

DFT calculations solve quantum mechanical equations to provide:

Accurate binding energies
Electronic structure (band gaps, conductivity)
Partial atomic charges
Reaction energetics for catalysis

Multi-Scale Approach

flowchart TD A[DFT
Electronic structure
Accurate energies] --> B[Force Field
Development] B --> C[GCMC/MD
Adsorption, diffusion
Larger systems] C --> D[ML Models
Rapid screening
Millions of MOFs] D --> E[Experimental
Validation] style A fill:#e74c3c,stroke:#c0392b,color:#fff style C fill:#f39c12,stroke:#e67e22,color:#fff style D fill:#27ae60,stroke:#1e8449,color:#fff

5.6 Software and Tools

Structure Analysis

Tool	Purpose	Access
Zeo++	Pore analysis, surface area calculation	Free, open source
PLATON	Crystal structure analysis	Free for academics
CrystalExplorer	Hirshfeld surface analysis	Free

Molecular Simulation

Tool	Purpose	Access
RASPA	GCMC and MD for adsorption	Free, open source
LAMMPS	General MD simulations	Free, open source
VASP/Gaussian	DFT calculations	Commercial
CP2K	DFT (free alternative)	Free, open source

Machine Learning

Tool	Purpose	Access
MOFid	MOF identification and featurization	Free
matminer	Materials feature extraction	Free, Python
CGCNN	Crystal graph neural networks	Free, PyTorch
MOFTransformer	Transformer model for MOFs	Free

5.7 Future Directions

Autonomous MOF Discovery

The future points toward self-driving laboratories where AI plans experiments, robots execute synthesis, and results feed back to improve models.

flowchart LR A[AI Planner] --> B[Robotic
Synthesis] B --> C[Automated
Characterization] C --> D[Data Analysis] D --> A style A fill:#9b59b6,stroke:#8e44ad,color:#fff style B fill:#3498db,stroke:#2980b9,color:#fff

Foundation Models for Materials

Large pre-trained models (like GPT for language) are being developed for materials science. These could enable:

Zero-shot property prediction
Natural language queries about MOFs
Automated literature synthesis
Multi-modal understanding (structure + text + spectra)

Integration with Experiment

Closing the loop between computation and experiment:

Active learning: AI selects most informative experiments
Bayesian optimization: Efficient exploration of synthesis space
Digital twins: Virtual models of experimental setups

Grand Challenges

Open Problems in Computational MOF Science

Synthesizability prediction: Can we predict which MOFs are actually makeable?
Long-term stability: Predicting degradation over months/years
Multi-objective optimization: Balancing competing properties
Transfer learning: Applying knowledge across different MOF families
Uncertainty quantification: Knowing when to trust predictions

Summary

Key Takeaways

MOF databases (CSD, CoRE MOF, QMOF) provide the foundation for computational discovery
Machine learning enables rapid property prediction for millions of MOFs
Feature engineering translates MOF structures into ML-ready representations
Generative AI can design new MOFs optimized for target properties
Molecular simulation (GCMC, MD, DFT) provides physical understanding
Multi-scale approaches combine accuracy and speed
Future directions include autonomous labs and foundation models
The combination of AI + experiment is accelerating MOF discovery

Check Your Understanding

Question 1

Why is the CoRE MOF database "computation-ready" while the raw CSD is not?

Click for Answer

Answer: The CoRE MOF database has been processed to remove solvent molecules from pores, assign partial atomic charges, and standardize file formats. Raw CSD structures often contain disordered atoms, solvent molecules, and missing hydrogens that must be cleaned before molecular simulations can be run reliably.

Question 2

What is the main advantage of using graph neural networks (GNNs) for MOF property prediction compared to traditional ML with hand-crafted features?

Click for Answer

Answer: GNNs can learn relevant features directly from the atomic connectivity graph rather than relying on pre-defined descriptors. This makes them more flexible and often more accurate, especially for properties that depend on local chemical environments that might not be captured by traditional geometric features.

Question 3

Why is synthesizability prediction one of the grand challenges in computational MOF discovery?

Click for Answer

Answer: A computationally designed MOF may be thermodynamically stable but kinetically inaccessible - meaning there's no practical synthesis route to make it. Synthesizability depends on complex factors including precursor availability, solubility, crystallization kinetics, and competing phases. These factors are difficult to model computationally, so many "optimal" predicted MOFs cannot actually be made.

Resources for Further Learning

Databases

CoRE MOF: github.com/gregchung/gregchung.github.io
QMOF: github.com/Andrew-S-Rosen/QMOF
CSD MOF Subset: ccdc.cam.ac.uk

Software

Zeo++: zeoplusplus.org
RASPA: github.com/iRASPA/RASPA2
matminer: hackingmaterials.github.io/matminer

Tutorials and Courses

Materials Project Workshop materials
MOF-related sessions at ACS meetings
Online courses on computational materials science