Determining Atomic Structure from Spectroscopy via an Active Learning Framework

1. Problem Statement

Determining atomic structures from spectroscopic data is a fundamental challenge in materials science. While spectroscopy (e.g., X-ray diffraction, XANES, EXAFS) provides critical information about atomic arrangements, solving the inverse problem (deducing structure from spectra) is inherently difficult due to:

• Computational Cost: Simulating spectra for complex materials (especially XANES and EXAFS) is computationally expensive, making iterative refinement intractable for traditional solvers.
• Non-Uniqueness (Ill-Posedness): Multiple distinct atomic structures can produce nearly identical spectra (the "phase problem"), particularly in disordered, amorphous, or nanocrystalline systems.
• Limitations of Existing Solvers:
  • Gradient-based methods (e.g., Rietveld refinement): Require analytical gradients (often unavailable for complex spectroscopies) and good initial guesses.
  • Probabilistic methods (e.g., Reverse Monte Carlo - RMC): Require thousands of simulations to converge, which is prohibitive for expensive spectroscopies.
  • Data-driven methods: Often rely on large pre-existing datasets (e.g., Materials Project) which may lack quantitative accuracy or cover specific stoichiometries/polymorphs.

2. Methodology: ActiveStructOpt

The authors introduce ActiveStructOpt, a framework that combines Active Learning, Graph Neural Networks (GNNs), and Bayesian Optimization to efficiently determine atomic structures with minimal simulations.

Core Components:

• Surrogate Modeling with GNNs:
  • Instead of running expensive physics-based simulations for every candidate, the framework trains a GNN to predict spectra from atomic structures.
  • Representation: Atoms are nodes, and interatomic distances (up to 6–8 Å) are edges. This captures medium-range structural information relevant to spectroscopy.
  • Architectures: Uses TorchMD-NET (trained from scratch) or pre-trained Orb-v3 models via MatterTune.
• Active Learning Loop:
  • The framework does not rely on large pre-computed datasets. Instead, it samples structure-spectra pairs "on the fly."
  • Process:
    1. Initialize with a small dataset (10 random structures).
    2. Train the GNN surrogate.
    3. Use Bayesian Optimization to select the next candidate structure.
    4. Run the expensive simulation only for the selected candidate.
    5. Add the new data to the training set and retrain the model.
• Optimization Strategy:
  • Acquisition Function: Uses a modified Lower Confidence Bound (LCB) balancing exploitation (minimizing Mean Squared Error between predicted and target spectra) and exploration (maximizing prediction uncertainty to improve the model).
  • Constraints: Enforces physical constraints (e.g., minimum interatomic distances based on the Materials Project) via soft, differentiable penalties (Lennard-Jones repulsive terms) to prevent unphysical structures.
  • Sampling: Supports both fixed unit cells (random perturbation) and variable unit cells (space-group-based sampling).
• Multi-Objective Capability: The framework can simultaneously optimize against multiple spectroscopies (e.g., XANES + EXAFS) or combine spectroscopic fit with energy minimization (DFT) to constrain the solution space.

3. Key Contributions

1. Data-Efficient Framework: ActiveStructOpt achieves high-accuracy structural determination with a fraction of the simulations required by traditional methods (e.g., RMC), making expensive spectroscopies like XANES/EXAFS viable for iterative refinement.
2. Elimination of Pre-computed Datasets: By using active learning to generate problem-specific data, the method avoids the inaccuracies and coverage gaps of general-purpose datasets.
3. Global Optimization: The framework is robust against poor initial guesses, capable of finding global minima even when starting from structures with little similarity to the target (unlike local gradient-based methods).
4. Multi-Modal Integration: It provides a unified framework for simultaneous refinement against multiple spectroscopies and energetic constraints, addressing the non-uniqueness of single-spectroscopy solutions.

4. Results and Benchmarks

The framework was benchmarked against Reverse Monte Carlo (RMC), DiffPy (Levenberg-Marquardt), and Gaussian Process-based Bayesian Optimization across 130+ test cases involving diverse materials (crystalline, amorphous, perovskites, lithium nickel oxide).

• X-ray Pair Distribution Function (PDF):
  • ActiveStructOpt consistently outperformed RMC and DiffPy.
  • In "High Temperature Perturbations" and "Lithium Nickel Oxide Delithiation" benchmarks, it achieved median Mean Squared Errors (MSE) significantly lower than baselines with only 200 simulation calls.
  • RMC required >10,000 simulations to match the performance of ActiveStructOpt in many cases.
• X-ray Absorption Spectroscopy (XANES/EXAFS):
  • Amorphous Carbon: Successfully refined a diamond structure to match an amorphous target EXAFS spectrum, a task where RMC failed to converge even with double the simulation budget.
  • LiNiO₂ Jahn-Teller Distortions: Successfully identified local distortions (splitting of Ni-O bonds) that were invisible to standard XRD but present in EXAFS/XANES.
• Multi-Spectroscopy & Energy Constraints:
  • Joint Optimization: Simultaneously fitting XANES and EXAFS for LiNiO₂ reduced ambiguity compared to single-spectroscopy fits.
  • Energy Minimization: Adding energy as an objective in the Materials Project polymorphs benchmark improved the structural match rate from 3% to 12% without degrading spectral fit, demonstrating that thermodynamic constraints help resolve the inverse problem's ill-posedness.

5. Significance and Conclusion

• Scalability: ActiveStructOpt offers a scalable path to automated, multi-modal materials characterization, particularly for complex and disordered systems where diffraction alone is insufficient.
• Addressing Ill-Posedness: The work highlights that while a single spectrum rarely guarantees a unique structure, combining complementary spectroscopies and energetic constraints within an active learning framework significantly narrows the solution space.
• Future Impact: As forward simulation models and foundation models improve, this framework provides a flexible, extensible approach to resolving complex atomic structures that were previously computationally intractable.

The source code for ActiveStructOpt is publicly available, facilitating adoption across the materials science community.