COFAP: A Universal Framework for COFs Adsorption Prediction through Designed Multi-Modal Extraction and Cross-Modal Synergy

This paper introduces COFAP, a universal deep learning framework that leverages multi-modal feature extraction and cross-modal attention to achieve state-of-the-art, gas-agnostic prediction of covalent organic framework (COF) adsorption performance, thereby enabling efficient high-throughput screening and application-specific prioritization of candidate materials.

The Gist

Imagine you are a master chef trying to find the perfect recipe for a new dish. You have a library containing 70,000 different potential ingredients (Covalent Organic Frameworks, or COFs). Your goal is to find the specific ingredients that will best capture a "guest" (like a gas molecule) and hold onto it, or perhaps separate two different guests from each other.

Traditionally, finding the best recipe meant:

1. Cooking every single dish (simulating every structure) to see how it tastes. This takes months or years.
2. Relying on specific taste tests (gas-specific data) that only work for one specific guest. If you want to test a different guest, you have to start all over again.

This paper introduces COFAP, a "Super Chef AI" that changes the game. Here is how it works, broken down into simple concepts:

1. The Problem: Too Many Choices, Too Slow

The world of COFs is like a massive, infinite Lego set. You can build millions of structures. Testing them one by one with traditional computer simulations is like trying to taste every single possible cake recipe in the world to find the best one. It's too slow and expensive.

2. The Solution: The "Three-Eyed" Detective (Multi-Modal Extraction)

Instead of looking at the Lego structure with just one pair of eyes, COFAP uses three different "senses" to understand the material at the same time. Think of it like a detective solving a crime by looking at the crime scene from three angles:

• Eye 1: The X-Ray Scanner (Sectional Planes)
  Imagine slicing a 3D cake into 9 different flat slices. COFAP looks at these 2D slices to see the pattern of the "sprinkles" (atoms) and the "frosting" (chemical bonds). It learns the visual shape and layout of the pores (the holes in the cake).
• Eye 2: The Map Maker (Persistent Homology)
  This eye doesn't care about the frosting; it cares about the tunnels. It draws a map of the tunnels and loops inside the structure. It asks: "Are there dead ends? Are the tunnels connected? Is it a maze or a straight hallway?" This helps it understand how gas molecules can travel through the material.
• Eye 3: The Chemist's Notebook (Bipartite Graph)
  This eye zooms out to look at the "building blocks" rather than every single atom. It groups the Lego pieces into "Linkers" (the bricks) and "Linkages" (the glue). It learns the chemical personality of the glue. Does it like to hug certain molecules?

3. The Brain: The "Cross-Modal Synergy" (Fusion)

Having three different views is great, but they need to talk to each other.

• Old way: Just averaging the three views (like mixing all your senses into a blurry soup).
• COFAP's way: It uses a Cross-Attention Mechanism. Imagine a conductor in an orchestra. The "Sectional Plane" (the visual shape) is the lead violinist. The other two senses (the map and the chemistry) are the backup musicians. The conductor listens to the lead violinist and asks the backup musicians, "Hey, does your map support what the violinist is playing? If so, play louder!"
  This allows the AI to combine the best parts of all three views without getting confused by the noise.

4. The Result: A Universal Predictor

Because COFAP learns the shape and chemistry of the material itself, it doesn't need to be retrained for every new gas.

• No "Guest-Specific" Crutches: Old models needed to know the "temperature" or "pressure" of a specific gas to make a guess. COFAP looks at the house (the material) and says, "This house has a door size and a floor texture that is perfect for any guest of this size."
• Speed: It can screen 70,000 materials in the time it takes to brew a cup of coffee, whereas traditional methods might take months.

5. The "Smart Filter" (Weight-Adjustable Prioritization)

Once COFAP ranks the best materials, researchers might have different goals:

• Goal A: "I need a material that is super cheap to regenerate (like a reusable sponge)."
• Goal B: "I need a material that holds the maximum amount of gas, even if it's hard to clean."

COFAP has a dial (a weight-adjustable system). You can turn the dial to prioritize "Regenerability" or "Capacity." It then instantly re-ranks the list to show you the top 10 materials that fit your specific needs.

6. The Discovery: The "Goldilocks Zone"

By analyzing the winners, the researchers found a secret pattern. The best materials for separating gases (like Methane and Hydrogen) aren't the biggest or the smallest. They are in a "Goldilocks Zone":

• Pores: Just the right size (not too big, not too small) to let one gas in but squeeze the other out.
• Surface Area: Enough to grab the gas, but not so much that it loses its selectivity.
• Chemistry: A specific mix of carbon and sulfur atoms that acts like a magnet for the target gas.

Summary

COFAP is like a super-intelligent, multi-sensory architect who can look at a blueprint, a map of the tunnels, and the chemical composition of the bricks all at once. It tells you exactly which building design will work best for your specific needs, without needing to build and test every single one. It turns a years-long search into a matter of hours, helping us find better materials for clean energy and pollution control much faster.

Technical Summary

1. Problem Statement

Covalent Organic Frameworks (COFs) are promising materials for gas adsorption and separation due to their tunable porosity and chemical functionality. However, the vast combinatorial design space of COFs makes experimental or brute-force simulation-based screening (e.g., Grand Canonical Monte Carlo, GCMC) computationally prohibitive.

Existing machine learning (ML) approaches face two critical limitations:

1. Reliance on Gas-Specific Descriptors: Many models depend on pre-calculated thermodynamic descriptors (e.g., Henry coefficients, adsorption heat) derived from GCMC simulations. These features are computationally expensive to generate, limit scalability, and reduce model transferability to new gases or operating conditions.
2. Incomplete Structural Representation: Alternative models that avoid gas-specific features often rely solely on simple geometric descriptors (e.g., from Zeo++) or single-modal representations. These fail to capture the complex interplay of 3D topology, chemical connectivity, and local atomic environments necessary for accurate structure-property mapping.

2. Methodology: The COFAP Framework

The authors propose COFAP, a universal, gas-agnostic framework that predicts adsorption performance using only the structural and chemical information inherent in the Crystallographic Information File (CIF). The framework consists of four main stages:

A. Multi-Modal Feature Extraction

To comprehensively capture the complexity of COFs, COFAP employs three distinct deep learning routes to extract complementary features:

1. Sectional Plane - Convolutional Variational AutoEncoder (SP-cVAE):
   • Concept: Reduces 3D COF structures into interpretable 2D representations.
   • Process: The COF supercell is sliced along nine representative crystallographic directions. Atoms (C, H, O, N) and bonds are projected onto 2D planes as dual-channel images (atom channel and bond channel).
   • Output: A Convolutional VAE compresses these 2D maps into compact latent descriptors, capturing global pore features, channel alignment, and chemical patterns.
2. Persistent Homology - Neural Network (PH-NN):
   • Concept: Captures 3D topological fingerprints and macrostructural parameters.
   • Process: Combines Persistent Homology (extracting H0 connectivity and H1 loop/tunnel features via Vietoris-Rips complexes) with standard geometric descriptors (Pore Limiting Diameter, Largest Cavity Diameter, Accessible Surface Area, density, porosity) calculated by Zeo++.
   • Output: Topological fingerprints that encode pore connectivity, nested relationships, and diffusion pathways, complementing the 2D plane data.
3. Bipartite Graph - Contrastive AutoEncoder (BiG-CAE):
   • Concept: Focuses on the core chemistry of linker-linkage interactions while eliminating atomic redundancy.
   • Process: Represents the COF as a coarse-grained bipartite supragraph where nodes are linkers (organic building blocks) and linkages (connection motifs like imine, amide). A contrastive autoencoder learns hidden group chemical features.
   • Output: Chemical descriptors representing the specific connectivity and functional groups governing host-guest interactions.

B. Cross-Modal Feature Fusion

• Mechanism: The framework uses a Cross-Attention Mechanism to fuse the features.
• Strategy: The SP-cVAE (providing the most comprehensive structural/chemical view) acts as the Query (Q). The PH-NN and BiG-CAE act as Keys (K) and Values (V).
• Benefit: This allows the model to selectively attend to relevant topological and chemical details from the auxiliary branches, creating a synergistic representation that is more robust than any single modality. All pre-trained encoder weights are frozen during fusion to preserve learned representations.

C. Weight-Adjustable Prioritization

• To facilitate practical screening, the authors developed a pipeline that converts predictions into a ranked list of candidates.
• It utilizes a linear weighted composite score combining Regenerability (R%) and Adsorbent Performance Score (APS).
• Researchers can adjust weights ($w_R, w_A$) to prioritize different application scenarios (e.g., high throughput vs. high selectivity) and perform sensitivity analysis to identify robust candidates.

3. Key Contributions

• Universal, Gas-Agnostic Prediction: COFAP achieves state-of-the-art performance without relying on gas-specific thermodynamic descriptors (like adsorption heat), enabling rapid screening for any gas without re-simulation.
• Novel Multi-Modal Architecture: The integration of 2D sectional projections, 3D topological fingerprints, and coarse-grained chemical graphs provides a holistic view of COF structures that single-modal models cannot achieve.
• Cross-Modal Synergy: The use of cross-attention effectively fuses complementary features, demonstrating that the whole is greater than the sum of its parts.
• Actionable Screening Pipeline: The framework includes a reproducible workflow for high-throughput screening, offering not just predictions but also structural statistics (e.g., optimal pore sizes, linker frequencies) and sensitivity analysis for decision-making.

4. Results

The framework was evaluated on the hypoCOFs dataset (69,840 computationally generated structures) for various tasks:

• Prediction Targets: Single-component gas uptake (CH4, H2, CO2, N2, O2 at various pressures) and CH4/H2 separation performance (Selectivity $S_{CH4/H2}$ and Working Capacity $\Delta N_{CH4}$).
• Performance Metrics:
  • Accuracy: COFAP achieved $R^2$ values exceeding 0.9 for most targets (e.g., $R^2 \approx 0.94$ for CH4/H2 selectivity under VSA conditions).
  • Comparison: It significantly outperformed existing ML models (Kernel Ridge, Random Forest, XGBoost) and models from literature that relied on gas-specific features. Notably, it surpassed models using adsorption heat inputs in selectivity tasks.
  • Efficiency: Inference speed averaged 158 samples/second on an RTX 4090 GPU, orders of magnitude faster than GCMC simulations.
• Ablation Studies: Removing any single modality (SP-cVAE, PH-NN, or BiG-CAE) resulted in performance degradation, confirming the necessity of the multi-modal fusion.
• Structural Insights: Statistical analysis of top-performing COFs revealed that high-performance candidates for CH4/H2 separation concentrate in narrow windows:
  • Pore Limiting Diameter (PLD): ~3.5–6.9 Å.
  • Largest Cavity Diameter (LCD): ~4.8–13.1 Å.
  • Specific Surface Area: Moderate ranges (avoiding excessively high values that dilute selectivity).
  • Chemistry: Overrepresentation of direct C–C linkages and specific linker motifs (e.g., imine groups) that enhance dispersive interactions.

5. Significance

• Accelerated Discovery: COFAP provides a practical route to screen tens of thousands of COF candidates in hours, drastically reducing the time and cost associated with traditional high-throughput screening.
• Interpretability: By identifying specific structural "windows" (PLD, LCD, porosity) and chemical motifs associated with high performance, the model offers actionable guidelines for the inverse design of new materials.
• Generalizability: The framework's ability to learn structure-property relationships without gas-specific inputs makes it a versatile tool applicable to various crystalline porous materials (e.g., MOFs, Zeolites) and diverse separation challenges.
• Foundation for Future Work: The study lays the groundwork for next-generation materials informatics, demonstrating how multi-modal deep learning can bridge the gap between complex crystal structures and functional performance.