🔬 atomic physics

A Machine-Learned Interatomic Potential Free-Energy Surface for the First Step of MIL-101(Cr) Secondary Building Unit Formation

This study employs a fine-tuned machine-learned interatomic potential (MACE-POLAR-M) to efficiently generate a converged free-energy surface for the initial step of MIL-101(Cr) secondary building unit formation, correcting the thermodynamic inaccuracies of a pre-trained model and revealing a previously unknown stepwise water-release mechanism that was previously inaccessible via computationally prohibitive ab initio molecular dynamics.

Original authors: Orlando A. Mendible-Barreto, Yamil J. Colón

Published 2026-10-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Orlando A. Mendible-Barreto, Yamil J. Colón

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where tiny, porous crystals act as microscopic sponges, capable of trapping gases, filtering water, or delivering medicine with surgical precision. These materials, known as metal-organic frameworks, are built from metal clusters connected by organic linkers, forming a vast, modular architecture. Scientists have long dreamed of designing these structures on a computer before ever mixing chemicals in a lab, hoping to predict exactly how they will form. However, the process of building these materials is incredibly complex. It involves atoms moving, bonds breaking, and new connections forming in a chaotic dance of energy that is too fast and too intricate for standard computer simulations to follow. To see these events clearly, researchers need a level of detail that usually requires supercomputers running for years, making the task of understanding how these materials assemble practically impossible with traditional methods.

A team of researchers at the University of Notre Dame has found a way to bypass this computational bottleneck, offering a new window into the very first steps of how a specific, highly useful crystal called MIL-101(Cr) begins to take shape. Instead of relying on the slow, expensive calculations that have historically blocked progress, they trained a specialized artificial intelligence model to act as a highly accurate, yet incredibly fast, substitute for those heavy computations. By teaching this AI model to recognize the subtle forces between atoms using a small set of high-quality reference data, the team was able to map out the energy landscape of the reaction. This map reveals not just where the atoms want to go, but the exact path they take and the energy hurdles they must overcome to build the material.

The study focused on the moment a chromium ion, surrounded by water molecules, begins to grab onto an organic linker to start forming the crystal's core. Previous computer models had suggested a specific path for this assembly, but they lacked the precision to confirm the energy changes involved. The researchers first tested a pre-trained AI model, one that had learned general chemistry from a vast database but had never seen this specific reaction. While this model could watch the atoms move, it got the thermodynamics completely wrong. It predicted that the reaction would require an input of energy to proceed, suggesting the product was less stable than the starting materials, and it identified a chemically incorrect structure as the most stable state. In short, the generalist AI could see the actors on stage but did not understand the plot.

To fix this, the team fine-tuned the model using a compact set of highly accurate calculations specific to this reaction. This process corrected the model's understanding of the energy landscape. The refined AI predicted that the reaction releases energy, making the product stable, and correctly identified the final structure as the most favorable state. With this corrected model, the researchers ran long simulations to generate a detailed free-energy surface, a two-dimensional map showing the stability of every possible arrangement of atoms during the reaction. This map revealed a surprising detail about the mechanism: the formation of the crystal core does not happen in one smooth, simultaneous motion as some earlier theories suggested. Instead, the process occurs in two distinct steps. First, one water molecule leaves the chromium ion to make space for a single oxygen from the organic linker. Then, a second water molecule departs, allowing the linker to fully lock into place.

The researchers also discovered that the accuracy of their AI model depends entirely on what data they used to teach it. When they included isolated fragments of the molecules in the training data, the model's performance on the actual reaction path degraded, showing that trying to teach the model too many unrelated details can confuse it. By restricting the training data to the specific configurations found along the reaction path, they ensured the model remained highly accurate where it mattered most. This approach successfully reproduced the known mechanisms of the reaction while refining the details of how the atoms move. The work demonstrates that machine learning, when carefully guided by high-quality data, can now calculate the energy surfaces of complex chemical reactions that were previously too expensive to study, opening the door to a deeper understanding of how these advanced materials are born.

Problem Statement
The self-assembly mechanisms of Metal-Organic Frameworks (MOFs), specifically the nucleation and growth of Secondary Building Units (SBUs), remain poorly characterized. While ab initio molecular dynamics (AIMD) offers the necessary accuracy to describe bond-forming and bond-breaking events, it is computationally prohibitive for converging the free-energy surfaces (FES) required to map these rare events over relevant timescales. Conversely, classical force fields lack the electronic structure detail needed for accurate chemical reactivity. This gap prevents a predictive understanding of MOF formation, leaving synthesis conditions largely dependent on empirical trial and error. Specifically, the first step of MIL-101(Cr) SBU formation—where a terephthalate (BDC) linker displaces water molecules from a chromium(III) center—lacks a converged, thermodynamically accurate FES at a high level of theory.

Methodology
The authors employed a six-stage computational pipeline to investigate the reaction: $[Cr(H_2O)_6]^{3+} + BDC^- \rightarrow [Cr(H_2O)_4–BDC]^{2+} + 2 H_2O$ .

Model Selection: The study utilized MACE-POLAR-M, a long-range-aware, equivariant machine-learned interatomic potential (MLIP) pre-trained on a broad chemical space.
Exploratory Sampling: Unbiased molecular dynamics (3 ns) and well-tempered metadynamics (7 ns) were performed using the pre-trained model at 300 K. Collective variables (CVs) included the coordination numbers between the Cr atom and water molecules, and between the Cr atom and BDC oxygens.
Reference Data Generation: Representative configurations from the metadynamics trajectory (spanning reactant, intermediate, and product basins) were selected for single-point Density Functional Theory (DFT) calculations. These were computed at the $\omega$ B97M-V/def2-TZVPP level of theory using ORCA, matching the training level of the foundation model.
Model Fine-Tuning: The pre-trained MACE-POLAR-M model was fine-tuned on this compact DFT dataset using three strategies:
- Naïve fine-tuning: Unrestricted updates to all weights.
- Low-Rank Adaptation (LoRA): Frozen weights with trainable low-rank update matrices.
- Multi-head (Replay) fine-tuning: Fine-tuning alongside a replay subset of the original foundation training data to mitigate catastrophic forgetting.
- Dataset Variation: Models were tested with and without dissociated fragment configurations (isolated ions and linkers) in the training set.
Production FES Calculation: Independent production metadynamics runs (7 ns) were conducted using the pre-trained and fine-tuned models to generate converged 2D free-energy surfaces.

Key Results

Pre-trained Model Limitations: The pre-trained MACE-POLAR-M model successfully sampled relevant configurations but failed thermodynamically. It predicted an endothermic reaction and identified a chemically incorrect global minimum (a Cr center coordinated to five water molecules with no linker), rather than the expected product basin.
Fine-Tuning Success: Both Naïve and Multi-head fine-tuned models corrected the thermodynamics, predicting an exothermic reaction and recovering the correct product basin (Cr coordinated to four waters and a bidentate BDC linker). These models achieved high accuracy on the test set (Energy MAE $\approx$ 1.5 meV/atom; Force MAE $\approx$ 30–40 meV/Å).
Mechanistic Refinement: The converged FES revealed a stepwise mechanism for water release, contradicting the previously proposed concerted single-step mechanism. The process involves two sequential energy barriers:
1. Release of the first water molecule to accommodate one BDC oxygen (forming an intermediate).
2. Release of the second water molecule to allow the formation of the complete bidentate configuration.
Training Data Sensitivity: The study found that including dissociated fragment configurations in the fine-tuning dataset degraded the model's accuracy on the specific reaction coordinate due to catastrophic forgetting of the general-purpose parametrization. Models trained without fragments maintained superior accuracy for the FES prediction task.
Barrier Discrepancy: The calculated free-energy barriers (approx. 13.5–14.2 kcal/mol for the first step) were significantly higher than the potential energy barrier (3.7 kcal/mol) reported in prior literature using PBE functionals and CI-NEB. The authors attribute this to the use of the more accurate $\omega$ B97M-V functional, the difference between potential energy (0 K) and free energy (finite temperature), and the methodological difference between minimum-energy path (CI-NEB) and enhanced sampling (metadynamics).

Significance and Claims
The paper establishes a validated, computationally tractable route to obtain free-energy surfaces for MOF formation chemistry at a level of theory ( $\omega$ B97M-V) where AIMD would be prohibitively expensive. The work demonstrates that fine-tuning foundation MLIPs on compact, reaction-specific datasets can correct thermodynamic errors and reveal mechanistic details (such as stepwise vs. concerted pathways) that are inaccessible to standard AIMD.

Crucially, the authors identify training-data composition as the key determinant of which reaction properties a fine-tuned model can reliably describe. They conclude that for FES prediction tasks, restricting fine-tuning data to configurations relevant to the reaction path (excluding dissociated fragments) protects accuracy along the reaction coordinate, whereas including fragments can lead to generalization failures on the primary task. This approach offers a transferable strategy for studying other MOF formation reactions where AIMD remains infeasible.

More like this