DFT Accuracy on Crystal Structure Prediction with Machine Learning Interatomic Potentials

The paper introduces CSP-MACE-Å, a machine learning interatomic potential that decomposes total energy into intra- and intermolecular components to achieve DFT-level accuracy in crystal structure prediction while running orders of magnitude faster, thereby enabling more robust derisking of solid forms through extensive candidate evaluation and free energy calculations.

The Gist

Imagine you are a chef trying to find the perfect recipe for a new cake. You have millions of potential ingredient combinations (candidate structures), but you only have time to taste-test a few dozen. To do this efficiently, you need a way to quickly guess which recipes are "good" before you actually bake them.

In the world of drug development, the "cake" is a medicine molecule, and the "recipe" is how those molecules stack together in a crystal. This stacking is called Crystal Structure Prediction (CSP). Getting the stacking right is crucial because different stacks (polymorphs) can make a drug dissolve too fast, not dissolve at all, or even turn into a different form while sitting on a shelf.

For years, the "gold standard" for tasting these recipes has been a super-accurate but incredibly slow computer simulation called DFT (Density Functional Theory). It's like a master chef who can taste a cake and tell you exactly how it will taste, but it takes them days to analyze just one recipe. Because it's so slow, scientists can only check a tiny fraction of the millions of possible recipes.

This paper introduces a new tool called CSP-MACE-Å. Think of this as a super-fast AI apprentice that has been trained to mimic the master chef's taste but can do the work thousands of times faster.

Here is how the paper explains this new tool, broken down into simple concepts:

1. The Two-Part Recipe (Intra vs. Inter)

The authors realized that a crystal is made of two types of interactions:

• Intramolecular: How the atoms hold together inside a single molecule (like the ingredients inside a single cookie).
• Intermolecular: How the molecules stick to each other to form the crystal (like how cookies stack in a jar).

The old AI models tried to learn everything at once and got confused. The new CSP-MACE-Å splits the job into two specialized teams:

• Team 1 (The Cookie Maker): Uses a model trained on a massive library of single molecules to understand how the ingredients hold together.
• Team 2 (The Jar Stacker): This is the secret sauce. It is specifically trained to understand the subtle ways molecules stick together in a crystal. It combines three things:
  1. A base model for sticking.
  2. A mathematical formula for long-range "van der Waals" forces (the weak magnetic-like pull between molecules).
  3. A "Delta Model" (a correction layer). This is like a taste-tester who only looks at the mistakes the other two made and fixes them to match the Master Chef's (DFT) results.

2. The Taste Tests (The Results)

The authors put their new AI apprentice through three rigorous taste tests to see if it could replace the slow Master Chef.

• Test 1: The AstraZeneca Kitchen (19 Compounds)
  They took 19 real-world drug compounds and asked the AI to rank the best crystal structures.

  • The Result: The AI's energy rankings were almost identical to the slow Master Chef (DFT).
  • The Twist: When they added a "temperature factor" (calculating free energy, which accounts for how the molecules wiggle and vibrate), the AI got even better, correctly identifying the most stable crystal form in almost every case.

• Test 2: The Blind Taste Test (28 Compounds)
  They tested the AI on 28 compounds from seven previous "blind tests" (where scientists didn't know the answer beforehand).

  • The Result: The AI performed just as well as the best DFT methods, and significantly better than other existing AI models.

• Test 3: The "ROY" Challenge (The Trickiest Cake)
  There is a famous molecule called ROY that has 14 different crystal forms. It is notoriously difficult because the molecules are flexible and tricky. Most computer models get this wrong.

  • The Result: Because their AI had a specialized "Cookie Maker" team trained on high-level chemistry, it correctly identified the most stable form of ROY, whereas other models failed.

3. Predicting the Future (Temperature Stability)

Finally, they tested if the AI could predict how the "cake" changes as the oven gets hotter. Some drugs are stable at room temperature but melt or change form when heated.

• They tested 5 compounds over a range of temperatures (from freezing to very hot).
• The Result: The AI successfully predicted the general trends. For example, it correctly guessed that one drug form is stable when cold, but a different form takes over when it gets hot. While it didn't get the exact temperature switch point perfect in every single case, it captured the overall behavior much better than previous methods.

The Bottom Line

The paper claims that CSP-MACE-Å is a breakthrough because it is fast enough to check millions of recipes but accurate enough to trust the results.

Instead of waiting days to check 100 recipes with the Master Chef, this AI can check thousands of recipes in the time it takes to brew a cup of coffee, with results that are nearly as accurate as the Master Chef. This allows scientists to "de-risk" their drug development by ensuring they don't miss a better, more stable crystal form that would have been too expensive to find with the old, slow methods.

What the paper does not claim:

• It does not claim this tool is currently being used in hospitals or for treating patients.
• It does not claim this will immediately cure diseases.
• It focuses strictly on the prediction of crystal structures, not on the chemical synthesis or clinical trials of the drugs themselves.

Technical Summary

Technical Summary: DFT Accuracy on Crystal Structure Prediction with Machine Learning Interatomic Potentials

Problem Statement
Crystal Structure Prediction (CSP) is a critical component of drug development, essential for selecting the solid-state form of active pharmaceutical ingredients (APIs) to ensure bioavailability, manufacturability, and stability. The standard CSP workflow involves generating millions of candidate structures and ranking them to identify the most stable polymorphs. The current state-of-the-art for the ranking phase relies on dispersion-corrected Density Functional Theory (DFT-D), such as PBE with Neumann–Perrin corrections or B86bPBE-XDM. While accurate, DFT is computationally expensive, often requiring hours for structure optimization and days for free energy calculations. This cost limits the number of candidate structures that can be evaluated and restricts the practical application of free energy reranking. Machine Learning Interatomic Potentials (MLIPs) offer a potential solution to accelerate these calculations by orders of magnitude, but previous attempts have struggled to match DFT accuracy, particularly in modeling long-range electrostatics, dispersion, and subtle intermolecular interactions in crystals.

Methodology: CSP-MACE-˚A
The authors present CSP-MACE-˚A, a specialized MLIP designed to replace DFT in CSP workflows. The model employs a decomposed energy architecture, separating the total energy into intramolecular and intermolecular components to allow for tailored modeling of each interaction type:

1. Energy Decomposition: The total energy is defined as $E_{total} = E_{intra} + E_{inter}$. The intramolecular energy ($E_{intra}$) is calculated as the sum of energies of constituent molecules isolated in vacuum, while the intermolecular energy ($E_{inter}$) is the residual difference between the full periodic system and the sum of isolated molecules.
2. Intramolecular Component: This component utilizes the MACE-POLAR architecture trained on the OMol25 dataset (100 million $\omega$B97M-V/def2-TZVPD DFT calculations). This high-level theory dataset ensures accurate modeling of intramolecular conformational energies, addressing issues like delocalization errors found in standard DFT functionals for flexible molecules.
3. Intermolecular Component: This is a hybrid model combining three terms to capture crystal-specific interactions:
   • MACE-POLAR Contribution: Provides a baseline for intermolecular interactions but lacks long-range dispersion.
   • Dispersion Term: A fixed-parameter term following the functional form of the XDM (Exchange-Hole Dipole Moment) correction. Parameters are averaged from an internal set of 50,000 B86bPBE-XDM calculations.
   • Learned Delta Model: A neural network trained to predict the residual between B86bPBE-XDM DFT targets and the sum of the MACE-POLAR and dispersion terms. This model is trained on 50,000 B86bPBE-XDM crystal structure calculations. Crucially, the training targets are isolated intermolecular residuals, preventing the loss function from being dominated by the larger intramolecular energy signals.

Key Contributions

• Architecture: The development of a decomposed MLIP that explicitly separates intra- and intermolecular modeling, allowing the use of high-level theory data for intramolecular terms and targeted crystal data for intermolecular terms.
• Training Strategy: The implementation of a "delta learning" approach for intermolecular interactions, where the model learns the correction to a physics-based dispersion model and a foundational MLIP, specifically targeting B86bPBE-XDM accuracy.
• Comprehensive Evaluation: A rigorous benchmarking of CSP-MACE-˚A against existing foundation models (MACE-POLAR-1, UMA-OMC) and DFT standards across multiple datasets, including AstraZeneca's internal CSP publications and the seven CSP blind tests.

Results
The evaluation was conducted on three primary sets:

1. AstraZeneca (AZ) Set (19 compounds): CSP-MACE-˚A achieved performance comparable to PBE-NP DFT when ranking by energy. However, when structures were reranked using harmonic free energies (calculated at 300 K), CSP-MACE-˚A significantly outperformed energy-only ranking, placing experimental matches within the top 10 structures and within 0.36 kJ/mol of the minimum energy on average. It consistently outperformed both MACE-POLAR-1 and the UMA-OMC foundation model.
2. Blind Test Set (28 compounds): CSP-MACE-˚A demonstrated performance close to B86bPBE-XDM DFT. While energy-only ranking was marginally worse than B86bPBE-XDM, the inclusion of harmonic free energy reranking allowed CSP-MACE-˚A to surpass the DFT energy ranking performance, achieving an average rank of 2.96 compared to 3.25 for B86bPBE-XDM.
3. ROY (Red Orange Yellow) Case Study: ROY is a challenging molecule with 14 known polymorphs where standard DFT often fails due to intramolecular delocalization errors. CSP-MACE-˚A correctly predicted the experimentally most stable Form Y as the global minimum (within 0.5 kJ/mol), whereas B86bPBE-XDM placed it >5 kJ/mol higher. This success is attributed to the high-fidelity intramolecular model (trained on $\omega$B97M-V) combined with accurate intermolecular modeling.
4. Thermodynamic Stability: On a set of five compounds with known temperature-dependent polymorphic stability, CSP-MACE-˚A successfully captured broad trends in relative stability using harmonic free energy approximations. While it did not always reproduce the exact ordering of polymorphs across all temperatures, it correctly identified monotropic and enantiotropic relationships and transition points for several compounds (e.g., Mexiletine Hydrochloride, AZD5462).

Significance and Claims
The paper claims that CSP-MACE-˚A represents a significant step forward in making MLIPs viable for industrial CSP applications. By achieving DFT-level accuracy while running multiple orders of magnitude faster, the model enables the evaluation of a far larger pool of candidate structures than is currently feasible with DFT. This capability allows for more robust derisking of solid forms by reducing the likelihood that viable polymorphs are excluded from the ranking stage due to computational cost. Furthermore, the ability to perform free energy reranking on large datasets makes the assessment of temperature-dependent stability practical for a broader range of compounds. The authors conclude that while the current work maintains standard CSP workflows, the speed of CSP-MACE-˚A opens the door for future adaptations of the workflow itself, such as integrating MLIPs into the structure generation phase.