Reliable and Efficient Automated Transition-State Searches with Machine-Learned Interatomic Potentials

This paper demonstrates that hybrid workflows combining machine-learned interatomic potentials (particularly MACE-OMol25) with transition-state search algorithms can achieve near-DFT accuracy for diverse chemical reactions while reducing computational costs by up to 96% compared to conventional methods.

The Gist

Imagine you are a chef trying to invent a new recipe. You know the ingredients (reactants) and the final dish (products), but you don't know the secret "peak" moment in the cooking process where the magic happens. In chemistry, this peak moment is called the Transition State. It's the most unstable, high-energy point where old bonds break and new ones form.

Finding this moment is crucial for understanding how reactions work, but it's incredibly hard to do. Traditionally, chemists use a super-accurate but slow computer program (called DFT) to simulate this. It's like trying to find a needle in a haystack by examining every single piece of hay with a microscope. It's accurate, but it takes forever and costs a fortune in computer time.

This paper introduces a new, faster way to find these needles using Machine-Learned Potentials (MLIPs). Think of MLIPs as a "smart guesser" or a seasoned sous-chef who has tasted millions of dishes. They aren't as perfect as the microscope, but they are incredibly fast and usually very close to the truth.

Here is the breakdown of what the researchers did, using simple analogies:

1. The Problem: The Slow Hike

Finding a transition state is like trying to hike to the very top of a foggy mountain peak.

• The Old Way (DFT): You take a step, stop, pull out a super-accurate GPS, wait for it to calculate your exact location, take another step, and repeat. It's precise, but you only take a few steps a day.
• The Goal: We need to get to the top (find the transition state) quickly so we can map out thousands of mountains (discover new catalysts and materials).

2. The Solution: The "Smart Hiker" (MLIPs)

The researchers tested six different "Smart Hikers" (Machine-Learned Potentials). These are AI models trained on massive databases of chemical data.

• The Strategy: Instead of using the slow GPS for every step, they let the Smart Hiker run up the mountain first to get close to the peak. Then, they use the slow, accurate GPS just for the final few steps to confirm they are exactly at the top.
• The Algorithms: They tested two ways to climb:
  • CI-NEB: Like laying down a rope ladder between the bottom and the top and climbing it rung by rung.
  • FSM (Freezing-String Method): Like throwing a rope from both the bottom and the top and letting them meet in the middle. The researchers found this "rope meeting" method (FSM) was much better at finding the peak without getting lost.

3. The Results: Who Was the Best Hiker?

The team tested these methods on 58 different chemical reactions, ranging from simple organic molecules (like small hydrocarbons) to complex metal catalysts (used in making drugs or fuels).

• The Winner: The AI model trained on a specific dataset called OMol25 (specifically MACE-OMol25) was the champion.

  • Success Rate: It found the correct peak 96.6% of the time.
  • Speed: It reduced the number of times they needed to use the slow, expensive GPS (DFT) by 94–96%.
  • Analogy: If the old way took 100 hours to find a reaction path, this new way takes about 4 hours.

• The Runner-Up: Another model called UMA-Medium was fantastic for the complex "metal" mountains (transition metals), showing that these AI tools can handle difficult, heavy-duty chemistry too.

4. The Secret Sauce: "Low-Level Refinement"

One of the paper's biggest discoveries was a specific trick they used.

• The Trick: Before calling the expensive GPS, they let the Smart Hiker take a few extra steps to get really close to the peak on their own.
• The Benefit: This "low-level refinement" made the final GPS check much faster. It's like the Smart Hiker getting you to the base camp of the peak, so the GPS only has to do the final 10 meters of work instead of the whole mountain. This cut the computer cost by three times without losing accuracy.

5. Why This Matters

• For Chemists: This is a game-changer. It means we can now screen thousands of potential new drugs, batteries, or industrial catalysts in a fraction of the time it used to take.
• For the Future: While the AI is amazing for simple molecules, it still needs a little help with the most complex metal reactions. However, the paper proves that we are ready to use these tools for high-speed discovery.

The Bottom Line

This paper shows that we no longer have to choose between accuracy and speed. By combining a fast, smart AI guesser with a precise final check, chemists can now explore the chemical world at a speed that was previously impossible. It's like upgrading from a bicycle to a supersonic jet for discovering new materials.

Technical Summary

1. Problem Statement

Determining reaction mechanisms and transition states (TS) is critical for catalyst design and materials discovery. However, traditional methods relying on Density Functional Theory (DFT) are computationally prohibitive for high-throughput screening due to the non-linear scaling of DFT and the large number of potential energy surface (PES) evaluations required by TS-search algorithms (e.g., Nudged Elastic Band, Freezing String Method). While Machine-Learned Interatomic Potentials (MLIPs) offer near-DFT accuracy at orders-of-magnitude lower cost, their reliability for automated TS searches—particularly in locating the correct saddle point and handling diverse chemical spaces (organics, polymers, transition metals)—remains under-explored.

2. Methodology

The authors developed and benchmarked a hybrid two-stage workflow combining MLIPs with DFT:

• Stage 1 (Low-Level Search): Reaction path finding and initial TS guessing are performed using MLIPs or semi-empirical DFT. Two algorithms were tested:
  • Freezing String Method (FSM): Uses a "freeze-and-grow" strategy to incrementally extend strings from reactants/products, freezing nodes to reduce optimization costs.
  • Climbing-Image Nudged Elastic Band (CI-NEB): Optimizes a chain of states to find the minimum energy path (MEP).
• Stage 2 (Refinement & Validation): The candidate TS geometry is refined using high-level DFT ($\omega$B97X-V/def2-TZVP) via Partitioned Rational Function Optimization (P-RFO).
  • Native Workflow: The MLIP guess is passed directly to DFT refinement.
  • Low-Level Refined Workflow: The MLIP guess undergoes an additional local refinement step on the MLIP surface (using the Sella package) before DFT refinement.

Models Benchmarked:
Six potentials were evaluated across 58 diverse reactions:

1. MACE-OMol25 (Trained on Open Molecules 2025 dataset).
2. UMA-Medium & UMA-Small (Mixture-of-experts variants trained on OMol25 + other datasets).
3. eSEN-S (Equivariant graph transformer trained on OMol25).
4. AIMNet2 (Atoms-in-molecules neural network).
5. GFN2-xTB (Semi-empirical DFT).

Benchmarks:

• Organic: Baker set (24 reactions) and Sharada set (9 reactions).
• Polymerization: Poly25 set (25 closed-shell reactions).
• Organometallic: Four case studies including Scandium-catalyzed olefin insertion, Iron-catalyzed transfer hydrogenation, Platinum-mediated metallabenzene rearrangement, and Rhodium-catalyzed C–H activation (both in-distribution and out-of-distribution).

3. Key Contributions

• Systematic Benchmarking: The first large-scale evaluation of six distinct MLIPs combined with two path-finding algorithms across 58 reactions spanning small organics, polymers, and transition-metal catalysis.
• Workflow Optimization: Demonstration that a low-level refinement step on the MLIP surface significantly reduces the number of expensive DFT gradient evaluations required for final convergence without sacrificing reliability.
• Dataset Impact: Identification that models trained on the Open Molecules 2025 (OMol25) dataset (specifically MACE-OMol25 and UMA-M) significantly outperform models trained on other datasets or semi-empirical methods.
• Algorithm Comparison: Quantitative evidence that the Freezing String Method (FSM) is more robust and efficient than CI-NEB for automated TS searches in this context.

4. Key Results

A. Performance on Organic Systems

• Success Rates: The FSM algorithm achieved success rates of 88.9% (Baker) and 89.8% (Sharada), significantly outperforming CI-NEB (70.8% and 62.0%, respectively).
• Model Superiority: Models trained on OMol25 (MACE-OMol25, UMA-M, UMA-S, eSEN-S) showed a clear advantage. MACE-OMol25 achieved the highest success rate of 96.6% on organic benchmarks.
• Cost Reduction:
  • Native Workflow: Required ~14.7 DFT gradient evaluations per successful search.
  • Low-Level Refined Workflow: Reduced cost to ~5.0 evaluations (a ~66% reduction).
  • MACE-OMol25 Specifics: Achieved a mean cost of only 3.8 DFT gradient evaluations per successful search (a 94–96% reduction compared to all-DFT workflows).

B. Performance on Polymerization (Poly25)

• MACE-OMol25 and UMA-M achieved 96.0% success rates.
• Low-level refinement reduced the average DFT cost from 22–31 evaluations to **5 evaluations**.
• Failures were primarily attributed to interpolation issues in the FSM algorithm (internal coordinate to Cartesian transformation) rather than potential inaccuracies.

C. Performance on Organometallic Systems

• UMA-Medium demonstrated superior transferability for transition-metal systems, including out-of-distribution Rhodium-catalyzed C–H activation. This is attributed to its training data including metal-surface and metal-adsorbate configurations (OC20/OC22).
• MACE-OMol25 performed well but showed slightly lower reliability on complex transition-metal systems compared to UMA-M.
• Accuracy: MLIP-refined geometries were highly accurate (RMSD < 0.04 Å for reactive cores) compared to DFT references, though energetic barriers sometimes showed systematic errors (e.g., overestimation by 5–8 kcal/mol).

D. Failure Mode Analysis

• Hessian Errors: A primary cause of failure in low-level refinement was the inaccuracy of the approximate Hessian generated by MLIPs. This could misrank imaginary frequencies, causing the optimizer to converge to an alternate (incorrect) saddle point or a second-order saddle point.
• Convergence: Some failures occurred when the initial guess lay outside the basin of attraction of the target TS.
• Algorithm Sensitivity: The FSM's "freeze" strategy sacrifices the full MEP, which can lead to suboptimal initial guesses for complex pathways, though this was generally outweighed by its efficiency.

5. Significance and Conclusion

This study establishes MLIP-accelerated workflows as practical, high-throughput tools for automated reaction discovery:

1. Efficiency: By offloading the expensive reaction-path finding to MLIPs and using low-level refinement, the computational cost of TS searches is reduced by one to two orders of magnitude (94–96% reduction in DFT calls) while maintaining near-DFT reliability.
2. Scalability: The approach is scalable to large systems (polymers, organometallics) where pure DFT searches are infeasible.
3. Recommendation: The authors recommend adopting MACE-OMol25 (for small organics) and UMA-Medium (for transition metals) combined with the FSM algorithm and low-level refinement as the default first stage in TS-search workflows.
4. Future Outlook: While highly effective for organic and closed-shell systems, challenges remain for transition-metal systems involving complex spin/charge reorganization. Future improvements should focus on better Hessian estimation (via automatic differentiation) and MLIPs capable of handling variable electronic states.

In summary, the paper demonstrates that the combination of modern MLIPs (specifically those trained on OMol25) and efficient path-finding algorithms has matured to a point where automated, high-throughput transition-state searches are no longer limited by computational cost.