The combination of ML and AM for MOFs represents a frontier in functional materials manufacturing. ML enables rapid screening of the vast MOF chemical space, while AM provides pathways to manufacture complex geometries that maximize mass transport and functional performance.

Structure-Property Prediction

ML models trained on crystallographic data predict MOF properties orders of magnitude faster than molecular simulations.

Property	ML Method	Input Features	Accuracy
Surface Area (BET)	Random Forest, GNN	Topology, pore geometry	R2 > 0.95
Gas Uptake (CO2, CH4, H2)	Neural Networks	Chemical descriptors, PLD, LCD	R2 = 0.90-0.95
Thermal Stability	Gradient Boosting	Metal node, linker chemistry	85% classification
Water Stability	SVM, RF	Metal-linker bond strength	82% classification
Selectivity (CO2/N2)	GNN, Transformer	Pore environment	R2 = 0.88

Inverse Design

Generative models (VAE, GAN) design novel MOF structures with target properties.

• Crystal VAE: Generates valid MOF structures in latent space
• Reinforcement Learning: Optimizes linker-node combinations
• Diffusion Models: Emerging approach for crystal generation

Direct Ink Writing (DIW)

The most common AM method for MOFs, using paste formulations extruded through nozzles.

• MOF loading: Typically 50-80 wt% in binder matrix
• Binders: PVA, cellulose, bentonite clay
• Challenge: Maintaining pore accessibility after binding
• ML optimization: Ink rheology, print parameters, drying

Stereolithography (SLA)

• Approach: MOF particles suspended in photopolymer
• Resolution: Higher than DIW (50-100 microns)
• Challenge: Light scattering from MOF particles
• ML role: Cure depth prediction, particle distribution

MOF-on-Scaffold

• Approach: Print scaffold first, then grow MOF in situ
• Advantage: Maximizes MOF crystallinity and surface area
• Methods: Hydrothermal, solvothermal on 3D printed substrates

AM Method	Resolution	MOF Loading	Surface Area Retention
DIW (Robocasting)	200-500 microns	50-80 wt%	60-85%
SLA/DLP	50-100 microns	20-40 wt%	50-70%
MOF-on-Scaffold	Variable	N/A (grown)	90-100%
Binder Jetting	100-200 microns	60-90 wt%	70-80%

Graph Neural Networks for MOFs

GNNs treat MOF structures as graphs (atoms as nodes, bonds as edges) enabling direct learning from crystal structures.

• CGCNN: Crystal Graph Convolutional Neural Network
• MEGNet: Materials Graph Network
• MOFNet: Purpose-built for MOF property prediction

Geometric Descriptors

• PLD: Pore Limiting Diameter
• LCD: Largest Cavity Diameter
• VSA: Volumetric Surface Area
• VF: Void Fraction

Gas Storage & Separation

• CO2 capture: 3D printed MOF monoliths for direct air capture
• H2 storage: Optimized geometries for fuel cell applications
• Natural gas purification: CH4/CO2 separation

Catalysis

• Structured reactors: 3D printed catalyst beds with MOF coatings
• Photocatalysis: Light-harvesting MOF architectures
• Electrocatalysis: MOF-derived carbon materials

Sensing & Biomedical

• Gas sensors: 3D printed MOF sensor arrays
• Drug delivery: Controlled release from porous structures
• Water harvesting: Atmospheric moisture capture

View all 595 capstone papers →