Data-Driven Thermal and Mechanical Modeling of Defective Covalent Organic Frameworks

This study develops and validates a specialized machine learning interatomic potential (QCOF) based on the MACE architecture to efficiently simulate the thermal and mechanical properties of defective covalent organic frameworks, revealing distinct defect sensitivities in CTF-1 and COF-LZU1 systems while establishing a robust framework for large-scale quantum-accurate modeling of extended network materials.

The Gist

The Big Picture: Building Better "Smart" Lego

Imagine Covalent Organic Frameworks (COFs) as incredibly intricate, 2D sheets of molecular Lego. These aren't just any Lego; they are made of organic materials (like carbon and nitrogen) and are designed to be super-light, porous, and tunable. Scientists love them because they could be the next big thing in flexible electronics, gas storage, or sensors.

However, there's a catch. Just like real-world construction, you can't build these perfect Lego sheets without making a few mistakes. Sometimes a piece is missing, sometimes two pieces are glued together the wrong way, or a whole section is shifted. These are called defects.

The problem is that these tiny mistakes can ruin the sheet's performance. They might stop heat from flowing (like a clogged pipe) or make the sheet snap under pressure. To understand how to fix this, scientists need to simulate these sheets on a computer. But here's the dilemma:

• The "Gold Standard" (DFT): This is like trying to calculate the physics of every single atom using a super-accurate but incredibly slow calculator. It's so slow you can only simulate a tiny handful of atoms (a few hundred).
• The "Fast Approximation" (Classical Force Fields): This is like using a quick, rough estimate. It's fast enough to simulate millions of atoms, but it's often inaccurate and can't handle complex chemical changes (like breaking bonds).

The Goal: The authors wanted to build a "Goldilocks" tool: a computer model that is fast enough to simulate huge sheets (tens of thousands of atoms) but accurate enough to predict exactly how those defects change the material's heat and strength.

---

The Solution: The "Quantum COF" (QCOF) Model

The team created a new AI tool called QCOF (Quantum Covalent Organic Framework). Think of QCOF as a super-smart apprentice.

1. Training: They didn't just guess how the atoms behave. They trained the AI on a massive dataset of "perfect" and "stretched" Lego sheets, using the slow, super-accurate "Gold Standard" calculator to teach the AI the rules of physics.
2. The Result: The AI learned the rules so well that it can now predict how atoms will move and interact almost as accurately as the slow calculator, but it does it thousands of times faster.

They tested QCOF against other popular AI models (like MACE-OFF24) and found that QCOF was the winner. It was the most efficient, used the least computer memory, and was the most accurate at predicting how atoms push and pull on each other.

---

The Experiments: Heat and Strength

Once they had their super-tool, they put it to work on two specific types of COF sheets: CTF-1 and COF-LZU1. They introduced "defects" (like missing bricks or weirdly shaped holes) and watched what happened.

1. The Heat Test (Thermal Conductivity)

Imagine trying to run a relay race where the runners are heat energy (phonons) passing a baton.

• CTF-1 is like a rigid, stiff track. When you put a pothole (defect) in the track, the runners trip and the race slows down dramatically. The study found that CTF-1 is very sensitive to defects; even a small mistake drops its ability to conduct heat significantly.
• COF-LZU1 is like a bouncy, flexible trampoline. It's already wiggly and chaotic. Adding a pothole doesn't change much because the runners were already stumbling around anyway. This material is surprisingly resilient to defects when it comes to heat flow.

The Lesson: You can't assume all materials react to damage the same way. Stiff materials break easily when damaged; flexible materials are already "damaged" by their own flexibility, so extra damage doesn't hurt as much.

2. The Strength Test (Mechanical Stress)

Next, they pulled on the sheets like taffy to see how strong they were.

• The Stiffness (Young's Modulus): This is how hard it is to stretch the material. Surprisingly, adding a few defects didn't change the stiffness much. The sheet still felt "stiff" to the touch.
• The Breaking Point (Ultimate Strength): This is when the sheet snaps. Here, the defects mattered a lot. Even a single missing brick acted like a weak link in a chain. The sheet didn't get much "softer," but it became much more likely to snap suddenly under pressure.

The Lesson: A defective COF might still feel strong and rigid in normal use, but it's hiding a secret weakness. It might snap much sooner than a perfect sheet if you pull too hard.

---

Why This Matters

This paper is a breakthrough because it gives scientists a fast, reliable way to design better materials.

Before this, if you wanted to know how a defective COF would behave, you'd have to either:

1. Wait years for a supercomputer to calculate a tiny piece (too slow).
2. Guess using rough models that might be wrong (too risky).

Now, with QCOF, scientists can simulate massive sheets with thousands of atoms in a matter of days. This allows them to:

• Design better filters for gas storage that won't clog or break.
• Create flexible electronics that can bend without snapping.
• Predict failure points before building the material in a lab.

The Bottom Line

The authors built a "smart simulator" that bridges the gap between slow, perfect physics and fast, rough guesses. They used it to discover that stiff materials are fragile when damaged, while flexible materials are surprisingly tough. This insight helps engineers build the next generation of high-tech, defect-tolerant materials.

Technical Summary

1. Problem Statement

Covalent Organic Frameworks (COFs) are promising 2D materials for flexible electronics, catalysis, and sensing due to their tunable porosity and high surface area. However, real-world COFs inevitably contain synthesis-induced defects (e.g., missing linkers, Stone–Wales defects, grain boundaries) that significantly alter their intrinsic thermal and mechanical properties.

Current computational methods face a "accuracy-efficiency gap":

• Density Functional Theory (DFT): Offers quantum accuracy but is computationally prohibitive for the large system sizes (tens of thousands of atoms) required to model defects and thermal transport phenomena.
• Classical Force Fields: Efficient for large scales but lack the accuracy to describe complex bond breaking, formation, and specific interatomic interactions, often failing to capture the nuances of defective structures.
• General Machine Learning Interatomic Potentials (MLIPs): While promising, existing general-purpose models often struggle with transferability to unseen chemical environments or lack the computational efficiency required for large-scale simulations of specific material classes like COFs.

The challenge is to develop a computational framework that combines quantum-level accuracy with the computational efficiency necessary to simulate large-scale, defective COF systems to predict thermal conductivity and mechanical strength.

2. Methodology

The authors developed a specialized workflow centered on a new MLIP architecture:

• Model Architecture: They utilized the MACE (Many-body Atomic Cluster Expansion) architecture, specifically tuning it for COFs. The resulting model is named QCOF ("Quantum COF").
• Training Data: The model was trained on an extensive dataset of over 36,000 non-equilibrium conformations of 23 distinct carbon-nitride nanosheets. These structures were generated via ab initio molecular dynamics (AIMD) using DFT (PBE+D3 level), including equilibrium states and structures subjected to large tensile strains up to fracture.
• Hyperparameter Optimization: The authors systematically varied three key hyperparameters:
  • Descriptor dimensionality ($D$): 32 to 256.
  • Equivariance ($E$): Invariant ($E=0$) vs. Equivariant ($E=1$).
  • Cutoff radius ($r_c$): 3 to 7 Å.
  • Result: The optimal configuration was found to be an invariant model with a small descriptor dimension ($D=48$) and a short cutoff ($r_c=4$ Å), labeled 48-0-4.
• Benchmarking: The QCOF model was compared against:
  • General-purpose MACE models (MACE-OFF24, MACE-MPA-0).
  • Fine-tuned variants of these general models.
  • A Moment Tensor Potential (MTP) trained on the same dataset.
• Simulation Techniques:
  • Thermal Transport: Non-Equilibrium Molecular Dynamics (NEMD) to calculate thermal conductivity ($\kappa$) using the direct method, accounting for finite-size effects via extrapolation.
  • Mechanical Properties: Uniaxial tensile loading simulations to generate stress-strain curves and calculate Young's modulus and fracture stress.
  • Phonon Analysis: Frozen phonon approach to calculate phonon dispersions and group velocities.

3. Key Contributions

1. Development of QCOF: A specialized, invariant MLIP architecture optimized for COFs that achieves a superior trade-off between accuracy, memory footprint, and inference speed compared to general-purpose models.
2. Large-Scale Defect Simulation: The ability to simulate systems with >40,000 atoms (up to 140,000 atoms) containing various defects (Stone–Wales, 5-8-5, grain boundaries) with quantum accuracy, a scale previously inaccessible to DFT and difficult for other MLIPs.
3. Comprehensive Property Prediction: A unified framework capable of predicting both thermal conductivity and mechanical strength (stress-strain behavior) for pristine and defective COFs.
4. Defect Sensitivity Analysis: A systematic investigation revealing that the impact of defects on thermal and mechanical properties is highly dependent on the intrinsic stiffness and vibrational spectrum of the specific COF material, not just defect density.

4. Key Results

Model Performance & Benchmarking

• Accuracy: The QCOF model (48-0-4) achieved a force Root Mean Square Error (RMSE) of 25.4 meV/Å on validation data, outperforming fine-tuned general-purpose models.
• Efficiency: QCOF is significantly faster and more memory-efficient. It can simulate a 140,000-atom system on a single NVIDIA H100 GPU, requiring only ~6 days for 1 ns of simulation. In contrast, general-purpose models (like MACE-OFF24) failed to scale to these sizes or were significantly slower.
• Transferability:
  • QCOF generalized well to unseen chemical compositions (e.g., CTF-1 and COF-LZU1, where CTF-1 was in training but LZU1 was not).
  • It maintained high accuracy for defective structures, with force errors remaining low even for Stone–Wales and 5-8-5 defects.
  • Phonon dispersion calculations for C2N matched well with Density Functional Perturbation Theory (DFPT) reference data.

Thermal Transport Properties

• Size Effect: Thermal conductivity ($\kappa$) of C2N converged to ~88 W m⁻¹ K⁻¹ after extrapolating to infinite length, validating the model against previous DFT and MTP studies.
• Defect Impact:
  • CTF-1 (Stiff): Showed a strong sensitivity to defects. Introducing point defects reduced $\kappa$ by ~10%, and grain boundaries reduced it by ~30%.
  • COF-LZU1 (Flexible): Showed minimal sensitivity to defects; $\kappa$ remained nearly unchanged.
  • Insight: Stiffer frameworks (CTF-1) support longer phonon mean free paths, making them more susceptible to scattering by defects. Flexible frameworks (LZU1) already have short mean free paths due to intrinsic anharmonicity, so additional defects have a negligible effect.

Mechanical Properties

• Elasticity: The 2D Young's Modulus remained virtually invariant at low defect concentrations for both materials.
  • CTF-1: ~69 GPa.
  • COF-LZU1: ~12–15 GPa (anisotropic).
• Strength: Defects had a drastic impact on ultimate tensile strength. A single defect in a large supercell reduced the breaking stress by up to 40%.
• Behavior: CTF-1 exhibited linear elastic behavior up to high strains, while COF-LZU1 showed anisotropic behavior at large strains.

5. Significance and Implications

• Bridging the Gap: This work successfully bridges the gap between the accuracy of first-principles methods and the scalability of classical force fields, enabling "quantum-informed" simulations of macroscopic material properties.
• Design Guidelines: The study provides critical design rules for COF-based devices:
  • Mechanical Reliability: While small defects do not significantly degrade the stiffness (Young's modulus) of COF membranes, they act as severe stress concentrators that drastically reduce fracture strength. Therefore, preventing crack nucleation is more critical than maximizing crystallinity for mechanical reliability.
  • Thermal Management: For thermal applications, the choice of COF topology is crucial. Stiff frameworks are highly sensitive to defect density, whereas flexible frameworks are robust against thermal degradation caused by structural disorder.
• Future Directions: The authors suggest that the QCOF framework can be extended to include oxygen-containing COFs and van der Waals interactions (essential for stacked layers) by expanding the training dataset, leveraging the scalability of the MACE architecture.

In conclusion, the QCOF model represents a significant advancement in materials informatics, providing a robust, scalable, and accurate tool for the rational design of next-generation COF materials for electronics and energy applications.