Surrogate Stability Screening

Updated 22 May 2026

Surrogate Stability Screening is a technique employing data-driven models to approximate stability criteria, enabling fast evaluation of complex systems.
It integrates adaptive sampling, ranking, and refinement pipelines to select candidates for high-fidelity validation in fields like materials discovery and power systems.
This approach offers significant computational speedups while addressing challenges associated with high-dimensional and parameterized systems.

Surrogate stability screening refers to the systematic use of surrogate models—data-driven or analytically tractable approximations—to rapidly assess the stability of candidate systems, structures, or hypotheses in lieu of evaluating the full, high-fidelity or first-principles model at every point of interest. This practice is motivated by the prohibitive computational cost and data requirements of exhaustive stability calculations for high-dimensional, complex, or highly parameterized systems—across fields including materials discovery, structural dynamics, fluid mechanics, machine learning explainability, power systems, and Bayesian clinical trial design. Surrogate stability screening encompasses not only the task of approximating stability indicators but also the process of adaptively sampling, ranking, and selecting candidates for further simulation or validation, often embedded in discovery or optimization loops.

1. Mathematical Foundations and Formal Criteria

Surrogate stability screening is built upon the definition of a stability criterion pertinent to the domain. These criteria are encapsulated in scalar or vector outputs, which are predicted by surrogates:

Materials and Crystallography: Stability is typically quantified as thermodynamic formation energy or energy above the convex hull, e.g., $\Delta E_{\mathrm{hull}} \le 0.1\,\mathrm{eV/atom}$ . Surrogates may repurpose latent distance metrics, such as the Crys-JEPA energy-aware embedding distance, as proxies for formation energy (Liu et al., 14 May 2026).
Geotechnical Systems: Slope stability is binary (failure/no-failure) according to physical or simulation-derived criteria, with models classifying or regressing on the state based on field or material properties (Aminpour et al., 2022).
Dynamical Systems: Both linear and nonlinear stability domains are delineated using spectral criteria (e.g., real parts of eigenvalues, Lyapunov exponents), with surrogates emulating leading eigenvalue statistics or classifier boundaries around chaotic regimes (Sousedík et al., 2021, Fuhg et al., 2019, Pia et al., 26 Jun 2025).
Power Systems: Stability corresponds to the ability of a dynamical grid model to maintain frequency above a safe nadir post-contingency, encoded as a binary or probabilistic outcome predicted by a neural classifier (Garcia et al., 3 Feb 2025).
Neural Surrogates in Engineering: Local stability of a surrogate regression is formally specified as boundedness of output perturbations against admissible input noise, encoded as constraint violations over specified boxes in the input domain (Ducoffe et al., 2024).
Clinical Trials: Surrogate validity is operationalized as strong and stable regression relationships between biomarker and clinical effect sizes, evaluated globally and within clusters via Dirichlet process mixture regression (Sachs et al., 2022).

2. Surrogate Model Classes and Construction

A diverse range of surrogate model architectures is deployed, selected for their flexibility, tractability, or suitability to the underlying structure:

Energy-Aware Embeddings and Transformers: Crys-JEPA leverages a deep Transformer encoder with InfoNCE-based energy-structured losses to map crystals to a continuous embedding space, in which latent distances correlate with formation energy (Liu et al., 14 May 2026).
Gaussian Processes and Kriging: Both regression and classification GPs (ordinary Kriging, Matérn kernels) are used extensively for functions with small-to-moderate sampling budgets, offering predictive distributions and uncertainty estimates for stability boundaries (Sousedík et al., 2021, Fuhg et al., 2019, Pia et al., 26 Jun 2025).
Neural Network Classifiers: Deep feedforward NNs (with specialized architectures and regularization) are employed for high-dimensional or discontinuous binary classification landscapes, as in power grid frequency stability and multi-output regression in aircraft engineering (Garcia et al., 3 Feb 2025, Ducoffe et al., 2024).
Monte Carlo and Ensemble Descriptors: In slope stability, ensemble bagging (random forests, SVC, stacking) is preferred for classification on simulated random field parameter sets (Aminpour et al., 2022).
Analytical Surrogates: Electrostatic surface stability is screened using explicit Ewald-sum calculations and slab dipole descriptors instead of ab initio DFT methods (Baiju et al., 29 Apr 2026).
Bayesian Nonparametrics: Hierarchical Dirichlet process mixture modeling is used to flexibly capture cluster-wise surrogate–clinical relationships in adaptive platform trials (Sachs et al., 2022).

The screening procedure interleaves surrogate evaluation with candidate selection, ranking, and, often, adaptive refinement:

Embedding-Based Rankings: In Crys-JEPA, distance $D_C$ in learned latent space to known stable reference structures is computed and used to rank generated crystals, with only the top fraction selected for further refinement and retraining of the generator (Liu et al., 14 May 2026).
Classifier-Based Filtering: In geotechnical and dynamical contexts, high-throughput screening is performed by evaluating the classifier surrogate on large parameter grids, replacing nearly all Monte Carlo model runs with rapid predictions, except for a small validation set near key thresholds (Aminpour et al., 2022, Sousedík et al., 2021, Fuhg et al., 2019).
Active Sampling: Power system stability surrogates are iteratively improved by active data collection near the predicted stability boundary, using the constrained OPF solver to preferentially sample uncertain points and thus sharpen the surrogate’s discriminatory capability (Garcia et al., 3 Feb 2025).
Sequential Verification Pipelines: Verification of neural surrogates for local stability progresses through adversarial attack (to find concrete violations), incomplete formal bound computation (to confirm verification where possible), and complete MILP formulations (to prove or disprove stability exhaustively) (Ducoffe et al., 2024).
Cluster-Based Stability Assessment: In Bayesian surrogate evaluation, cluster assignments are updated via Gibbs sampling, and stability of the surrogate relationship is confirmed only if all relevant clusters maintain sufficient regression strength and low residual variance (Sachs et al., 2022).

4. Evaluation Metrics, Validation, and Error Characterization

Metrics are tailored to the stability screening context, balancing the need for high fidelity and computational efficiency:

Domain/Task	Key Surrogate Metric	Direct Metric/Validation Reference
Crystal Discovery	$V.S.U.N$ (valid, stable, unique, novel rate)	DFT-calculated $\Delta E_{\mathrm{hull}}$ (Liu et al., 14 May 2026)
Slope Stability	Classification ACC, AUC, $\lvert \Delta P_f \rvert$	Full MC pf calculation (Aminpour et al., 2022)
Fluid Dynamics	RMSE on rightmost eigenvalue, instability probability	Direct eigenproblem solutions (Sousedík et al., 2021, Pia et al., 26 Jun 2025)
Power Systems	Fraction of stable dispatch under random validation	Full dynamic simulation of outages (Garcia et al., 3 Feb 2025)
Surrogate Regression	Verified/Violated point tallies, pipeline runtime	MILP full verification (Ducoffe et al., 2024)
Clinical Surrogacy	Clusterwise slope/posterior, $R^2$ , predictive error $\hat D$	Simulated or held-out clusters (Sachs et al., 2022)

Validation against high-fidelity calculations, simulation, or experiment is standard. Typical reported surrogate screening errors are sub-1% on probability estimation (slope failure), RMSEs $10^{-6}$ – $10^{-8}$ (Navier–Stokes eigenvalues), or $\leq 2\%$ (unstable grid dispatches after surrogate-constraint OPF).

5. Domain-Specific Implementations and Case Studies

Materials Discovery: Crys-JEPA achieves an 81.4–82.6% improvement in $D_C$ 0 over strongest baselines, operating at $D_C$ 1 s for 10,000 candidates versus hundreds–thousands of seconds for DFT or force-field approaches (Liu et al., 14 May 2026).
Slope Reliability: ML surrogates trained on 0.4% of MC data accurately classify the remainder, reducing computational cost from ≥306 days to <6 hours on 120,000 cases (Aminpour et al., 2022).
Fluid Mechanics: For dynamical systems, data-driven surrogate normal forms allow rapid bifurcation analysis and stability tracing, yielding $D_C$ 2– $D_C$ 3 accelerations for parameter continuation in Navier–Stokes flows (Pia et al., 26 Jun 2025).
Power Systems: Neural surrogates embedded in AC-OPF formulations eliminate frequency instability (12.8% → 0%) for practical thresholds, with solve time increases remaining below an order of magnitude (Garcia et al., 3 Feb 2025).
Engineering Surrogates: Combined empirical/formal verification pipelines can formally certify the stability of high-dimensional regression surrogates over norm-bounded input perturbations, critical for aerospace certification (Ducoffe et al., 2024).
Bayesian Clinical Trials: The DP mixture approach detects clusters of surrogate validity and flags subgroups where predictive stability fails, improving upon standard parametric approaches in adaptive designs (Sachs et al., 2022).

6. Computational Efficiency, Scalability, and Limitations

The computational gains inherent in surrogate stability screening enable large-scale and high-throughput studies not otherwise tractable:

Efficiency: Orders-of-magnitude speedup (e.g., $D_C$ 4– $D_C$ 5) by replacing MC, DFT, or time-domain integrations with cheap, vectorized surrogate evaluations (Liu et al., 14 May 2026, Aminpour et al., 2022, Sousedík et al., 2021, Pia et al., 26 Jun 2025).
Scalability: Surrogate pipelines support screening tens of thousands to millions of candidates and adaptive refinement on demand (Baiju et al., 29 Apr 2026, Garcia et al., 3 Feb 2025, Fuhg et al., 2019).
Limitations: Surrogates may lack accuracy in extreme regimes (high anisotropy, strong turbulence, highly nonlinear relationships), are sensitive to training data representativeness, and may not capture structural or mechanistic subtleties outside their parameter manifold. Formal specification and certification remain open challenges in many domains (Ducoffe et al., 2024, Sachs et al., 2022).

7. Outlook and Methodological Innovations

Surrogate stability screening is witnessing rapid methodological progress, merging classical design of experiments, modern ML, formal verification, and domain-specific screening heuristics:

Advances in embedding-based surrogates are enabling physical-property-aware ranking and fine-grained discriminations (Liu et al., 14 May 2026).
Adaptive sampling and classifier-based boundary refinement (e.g., MiVor, active-OPF) efficiently concentrate expensive simulations at the stability front (Fuhg et al., 2019, Garcia et al., 3 Feb 2025).
Sequential verification with empirical, incomplete, and complete formal testing ensures scalability while certifying local stability in critical applications (Ducoffe et al., 2024).
Hierarchical and nonparametric surrogacy evaluation frameworks accommodate clusterwise surrogate breakdowns and heterogeneous populations in complex trial designs (Sachs et al., 2022).
Domain transfer—embedding surrogates trained in one chemical space or design context into others—remains an active area of computational materials and systems research.

Surrogate stability screening thus provides the essential computational toolkit for globally- or locally-constrained discovery, design, and optimization under stability requirements across scientific and engineering disciplines.