Papers
Topics
Authors
Recent
Search
2000 character limit reached

Crys-JEPA: Accelerating Crystal Discovery via Embedding Screening and Generative Refinement

Published 14 May 2026 in cs.LG | (2605.14759v1)

Abstract: De novo crystal generation seeks to discover materials that are not merely realistic, but also stable and novel. However, most existing generative models are trained to maximize the likelihood of observed crystals, which encourages samples to stay close to known materials yet not necessarily align with the criteria that matter in discovery. Through an empirical investigation, we show that current crystal generative models are caught in a pronounced stability--novelty trade-off: moving toward the observed distribution preserves stability but limits novelty, whereas moving away from it quickly destroys stability. This suggests that the useful region for discovering crystals that are both stable and novel is extremely narrow. To escape the trade-off, we introduce Crys-JEPA, a joint embedding predictive architecture for crystals that learns an energy-aware latent space preserving formation-energy differences. In this space, stability assessment can be reformulated as an embedding-based comparison against accessible training crystals, reducing the reliance on expensive energy evaluation and task-specific external references. Building on Crys-JEPA, we further develop a screening-and-refinement pipeline that identifies promising generated crystals and reintroduces them to refine the generative model. On MP-20 and Alex-MP-20 datasets, we achieve improvements over baselines up to 81.4% and 82.6% on V.S.U.N metric, respectively.

Summary

  • The paper introduces an energy-aware JEPA framework that encodes formation energy within a latent space to effectively screen for thermodynamically stable crystals.
  • The paper employs a surrogate embedding distance as a DFT proxy, achieving up to 82.6% improvement on the V.S.U.N. metric while drastically reducing computational overhead.
  • The paper demonstrates a scalable screening-and-refinement pipeline that integrates energy-aware embeddings, reducing screening wall-time by 7×–17× compared to MLFF-based methods.

Crys-JEPA: Energy-Aware Embedding and Screening for De Novo Crystal Discovery

Introduction and Problem Analysis

The paper addresses de novo crystal generation, aiming to discover materials that are not only physically plausible but also thermodynamically stable and genuinely novel with respect to existing databases. A major challenge is the "stability-novelty trade-off" inherent in current generative models trained on log-likelihood objectives: outputs close to observed data are typically stable but not novel, while novel structures tend to forfeit stability. The region in design space supporting both properties is empirically shown to be extremely narrow. This inherent limitation restricts practical breakthroughs in materials discovery and calls for architectural and algorithmic innovations that can simultaneously promote novelty and stability.

Crys-JEPA Framework: Architecture and Training

To address this challenge, the authors propose Crys-JEPA, an energy-aware Joint Embedding Predictive Architecture (JEPA) specifically designed for crystals. Crys-JEPA is pre-trained to organize crystals in a latent space where proximity reflects similarity in formation energies, effectively encoding a crucial physical property directly into the embedding geometry. Key design features include:

  • Representation Learning: Crystals are represented by concatenated vectors of atomic coordinates, element-type one-hots, and lattice parameters, resulting in a stack of atom-wise vectors per crystal. Lattices are parameterized through a decomposition that factors out rotations and isolates a minimal parameter set.
  • Energy-Aware JEPA Training: Crys-JEPA uses physically meaningful data augmentations (isometric translations and rotations of the unit cell) and an energy-weighted InfoNCE loss, so that crystals with similar formation energies cluster in latent space while energetically dissimilar entries are repelled. This is crucial for screening with respect to stability.
  • Surrogate Stability Screening: By embedding a generated crystal and comparing distances to reference embeddings from the training set (which are assumed to be stable), the Crys-JEPA distance serves as a proxy for the crystal’s likely stability without requiring DFT or external references, significantly reducing computational overhead.

Screening-and-Refinement Pipeline

Crys-JEPA is embedded into a practical generation loop:

  1. Pre-train a base generative model (e.g., DDPM-Transformer) on training crystals.
  2. Generate a large pool of candidates.
  3. Relax generated candidates using a machine-learned force field and filter for basic validity, uniqueness, and novelty (V.U.N).
  4. For each V.U.N. crystal, compute embedding distance in Crys-JEPA space to known crystals of the same chemical system.
  5. Rank and select top candidates (smallest distance), reintroducing them into the training dataset for fine-tuning the generator, iterating this process.

Unlike earlier approaches (e.g., GNoME, MatterSim), this pipeline circumvents direct DFT calculations for most candidates, enabling high-throughput screening while ensuring generated crystals are not only valid and novel but also likely stable.

Experimental Results

Extensive experiments are reported on the established MP-20 and the large-scale Alex-MP-20 datasets. Key numerical results and claims include:

  • Pronounced Stability-Novelty Trade-off in Baselines: Empirically validated via multiple methods (CDVAE, MatterGen, ADIT, etc.), with strong negative correlation between stability and novelty; existing methods cannot simultaneously optimize both.
  • Substantial Improvement from Crys-JEPA Pipeline: The proposed screening-and-refinement pipeline yields improvements up to 81.4% (MP-20) and 82.6% (Alex-MP-20) on the stringent V.S.U.N. metric as evaluated by DFT—surpassing all strong baselines.
  • Efficient and Effective Screening: Crys-JEPA demonstrates a ~7×–17× reduction in screening wall-time compared to MLFF-based methods, while using substantially fewer input features and only energy labels (rather than force/stress). Crys-JEPA embeddings also encode more structural and compositional information than scalar machine-learning surrogates.
  • Latent Space Visualizations: The learned embedding manifold is stratified by formation energy, empirically supporting the use of Euclidean embedding distance as a meaningful stability proxy.
  • Comparisons to Fingerprints and MLFFs: Structural fingerprints are not energy-aware and thus perform worse as stability surrogates, while MLFF-based methods, though strong, are computationally more expensive and require more supervision.

Theoretical and Practical Implications

The findings have immediate implications for materials informatics and generative scientific discovery:

  • Empirical Characterization of Feasible Design Space: The paper provides strong evidence that the region where stability and novelty overlap is extremely narrow and models trained only for data likelihood are suboptimal for discovery.
  • Surrogate Model Design: Learning an energy-structured latent metric enables dramatic scaling of generative pipelines by removing the computational bottleneck of DFT, making in-silico screening practical for large candidate spaces.
  • Generality of Approach: While demonstrated for crystals, the joint embedding and energy-aware refinement paradigm is extensible to other scientific domains where competing objectives delimit a narrow feasible region.
  • Future Model Scaling: Crys-JEPA currently uses only energy labels and a moderate-scale dataset. Expanding to more input modalities (force, stress, symmetry), richer data (e.g., more exotic chemistries), or more powerful backbones (spectral or equivariant Transformers) may further enhance screening accuracy and model robustness.

Limitations and Future Directions

  • Surrogate Validation: Despite significant cost reduction, Crys-JEPA screen is ultimately a proxy and candidates must still be validated with first-principles calculations before experimental realization.
  • Latent Collapse Risks: While the energy-aware InfoNCE loss successfully structures latent space, risks of representation collapse remain, which could undermine fine-grained comparison. The use of SIGReg for collapse-resistance presents a promising research direction.
  • Scaling to Larger and Noisier Datasets: The method's effectiveness with more open-ended or noisy data distributions, such as hypothetical or low-quality structures, remains to be systematically explored.

Conclusion

Crys-JEPA introduces a scalable, physically informed latent screening paradigm for de novo crystal generation, directly addressing the narrow intersection of stability and novelty in material discovery. By structuring the learning task around energy-aware embeddings and embedding-based refinement, the method achieves strong efficiency and substantial improvements in joint generation metrics over state-of-the-art baselines. The approach presents an extensible template for surrogate-driven scientific search in other domains constrained by similarly stringent multi-objective criteria.

For further technical evaluation or reproduction, see the detailed architectural and training specifics in the appendix of (2605.14759).

Paper to Video (Beta)

No one has generated a video about this paper yet.

Sign Up to Generate All Videos Subscribe on YouTube

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Sign Up to Generate

Explain it Like I'm 14

Overview

This paper is about teaching computers to invent brand‑new crystals (solid materials with repeating atomic patterns) that are both safe to make and useful. The authors found that many AI models either play it too safe—sticking close to known crystals—or explore too wildly and suggest crystals that would fall apart. They propose a new helper model, called Crys‑JEPA, that quickly judges which AI‑invented crystals are likely to be stable, so the main generator can learn from the best ideas and improve.

What questions did the researchers ask?

The paper focuses on three simple questions:

  • Can we make AI that invents crystals that are both stable and genuinely new?
  • Why do current models get stuck choosing between “stable but not new” and “new but unstable”?
  • Can we build a fast, reliable way to screen many AI‑generated crystals and keep only the promising ones, without using very slow physics simulations?

How did they do it?

Think of crystal design like cooking:

  • The “ingredients” are chemical elements (like Li, O, Si).
  • A “recipe” is how many of each element you use and how you arrange the atoms.
  • A “good dish” (crystal) is one that holds together (stable), isn’t just a copy of a known recipe (novel), is correctly formed (valid), and is different from other dishes you’ve already made (unique). The authors call these V.S.U.N: Validity, Stability, Uniqueness, Novelty.

The challenge: a tightrope between stability and novelty

Most generators are trained to imitate what they’ve seen. That makes outputs close to known crystals (more stable) but less original. If the model wanders too far into unknown territory, stability drops fast. The authors show this is a real and strong trade‑off in existing systems.

The key idea: Crys‑JEPA (a fast “taste tester”)

Running full physics calculations (called DFT) to check stability for each candidate is very slow. Instead, the authors build Crys‑JEPA, a model that turns each crystal into a point in a special “map” (an embedding). In this map:

  • Crystals with similar formation energies (a measure related to how likely a crystal is to exist) are placed close together.
  • Crystals with very different formation energies are placed far apart.

You can think of it like a “flavor map” for crystals: similar taste = close; different taste = far. This lets the model compare a new, invented crystal to known stable crystals quickly, using distances in the map instead of heavy physics.

They train this map carefully so it respects changes that don’t affect energy (like rotating or shifting the unit cell), and they emphasize pushing apart crystals with big energy differences. That way, the distances in the map are meaningful for judging stability.

The screening-and-refinement loop

Using Crys‑JEPA, they create a simple improve‑your‑model cycle:

  • Train a basic crystal generator on known data.
  • Generate lots of new crystals and relax them with a fast learned simulator.
  • For each candidate, compare it (in the Crys‑JEPA map) to known crystals made from the same “ingredients.” Closer usually means more likely to be stable.
  • Keep the most promising ones and fine‑tune the generator on them.

By feeding the generator its own best creations, it learns to produce outputs that are both stable and more novel.

What did they find, and why does it matter?

  • They confirmed a strong stability–novelty trade‑off in many published generators: staying near known data keeps stability high but limits new discoveries; drifting away increases novelty but quickly harms stability.
  • Crys‑JEPA creates an energy‑aware “map” where closeness reflects similar formation energy. That makes fast, reasonable stability screening possible without full DFT.
  • Their screening‑and‑refinement pipeline significantly boosts overall quality on two standard benchmarks (MP‑20 and Alex‑MP‑20). Using a strict metric called V.S.U.N (which combines Validity, Stability, Uniqueness, and Novelty), they report improvements of up to about 81–83% compared to strong baselines when checked with DFT.
  • It’s also efficient. Compared to machine‑learning force fields (other fast energy estimators), Crys‑JEPA screens large batches much faster and doesn’t require full energy calculations for every reference crystal. It uses only the crystal structures and the learned embedding distances.

Why this matters: Faster, more reliable screening means you can explore far more candidate materials. That speeds up the path to finding new crystals for batteries, solar cells, or catalysts—places where better materials can make a big difference.

Key terms in simple language

  • Formation energy: Imagine the “cost” to assemble a crystal from its pure elements. Lower cost usually means a more stable material.
  • Energy above hull: At each “recipe” ratio, there’s a best‑possible energy. How much higher your crystal’s energy is above that best is the “energy above hull.” Below a small threshold (often 0.1 eV per atom) is considered stable.
  • Embedding: A way to represent complex objects (like crystals) as points in a space so that similar objects are near each other.

Implications and potential impact

  • Practical speed‑up: Labs and companies can use Crys‑JEPA to sift through huge numbers of AI‑generated crystals quickly, focusing expensive physics checks on only the best candidates.
  • Better balance: The refine‑by‑screening loop helps break the “stable or novel, pick one” problem, making true discoveries more likely.
  • Extensible idea: The approach could be scaled with more data or adapted to other material properties beyond stability.

A note of caution: Crys‑JEPA is a smart shortcut, not a final judge. Before real‑world use, the best candidates still need high‑accuracy validation (like DFT and experiments). Even so, this work shows a promising path to much faster materials discovery.

View Paper Prompt  View All Prompts 

Knowledge Gaps

Knowledge gaps, limitations, and open questions

Below is a single, concrete list of what remains missing, uncertain, or unexplored in the paper, intended to guide actionable follow-up research.

  • Quantitative calibration of the surrogate: The paper provides qualitative evidence (UMAP) that Crys-JEPA embeddings reflect formation-energy differences, but lacks a quantitative, cross-system analysis (e.g., Spearman/Kendall correlations vs. ∣ΔEf∣| \Delta E_f |, ROC/PR for stability classification) to calibrate embedding distance thresholds against DFT AEA E with confidence intervals.
  • Surrogate-to-DFT precision/recall: No systematic evaluation of the precision/recall of the screening step versus full DFT is provided (only 1,000-sample comparisons); future work should measure false positives/negatives by running DFT on both selected and rejected candidates to characterize selection reliability.
  • Heuristic distance vs. phase diagram: The average Euclidean embedding distance to a training reference set (Step 4) ignores the convex-hull geometry; methods to approximate or learn a convex hull in embedding space—or directly regress AEA E with calibrated uncertainties—remain unexplored.
  • Reference-set bias and completeness: Using the training set as the sole reference may miss unobserved low-energy competing phases; experiments that vary the reference set (e.g., older/newer MP snapshots, alternate databases) are needed to assess stability misclassification risks arising from incomplete references.
  • Normalization across chemical systems: The Dc metric is not normalized for per-system embedding dispersion or reference density; this could bias selection toward well-represented chemistries. Investigate system-aware normalization or diversity-constrained selection to avoid overfitting to dense regions.
  • Robustness to relaxation choices: The pipeline relaxes generated structures via MLFFs (MatterSim), but JEPA is trained on DFT labels. The effect of relaxer choice (MLFF vs. DFT) on embeddings and selection ordering is not quantified; ablate pre- vs. post-relaxation and cross-relaxer consistency.
  • Out-of-distribution generalization: Generalization of Crys-JEPA to unseen chemical systems, elements, and composition families (including those underrepresented in MP-20/Alex-MP-20) is untested; perform cross-dataset (e.g., OQMD, ICSD-like) evaluations and hold-out element/system studies.
  • Scalability beyond small cells: The approach is trained/evaluated on crystals with ≤20 atoms per cell; assess scalability and performance for larger unit cells, supercells, high-entropy alloys, complex frameworks, and structures requiring larger periodic images.
  • Handling of real-world complexities: The model does not address partial occupancies, disorder, defects, charge states, magnetism/spin ordering, anisotropic/stress effects, or van der Waals layering; extending embeddings and screening to these cases remains open.
  • Finite-temperature/pressure stability: Stability is assessed at 0 K using formation energies with a fixed AEA E threshold (0.1 eV/atom); incorporating free-energy contributions (phonons, configurational entropy), pressure effects, and metastability windows is an open direction.
  • Functional and dataset dependence: Crys-JEPA is trained on PBE-level data; transferability to other functionals (r2SCAN, HSE) or experimental stability labels is not evaluated. Systematic tests across functionals and re-calibration strategies are needed.
  • Iterative refinement dynamics: Only a single screening-and-refinement iteration is shown; the stability of multi-iteration loops (convergence, diversity retention, risk of mode collapse or overfitting to embedding-near regions) is unstudied.
  • Integration with stronger generators: The pipeline is demonstrated with a vanilla DDPM+Transformer; whether Crys-JEPA screening improves SOTA backbones (e.g., MatterGen, FlowMM, symmetry-aware diffusions) and by how much is not assessed.
  • Symmetry and invariance properties: The JEPA encoder uses a vanilla Transformer over concatenated coordinates/lattice without explicit permutation or space-group equivariance; the impact of incorporating symmetry-equivariant architectures on energy-aware embeddings remains untested.
  • Augmentation design in JEPA: Only rotation/translation augmentations are used; exploring small strains, symmetry-preserving perturbations, or contrastive positives that better reflect energy-preserving transformations could strengthen the embedding—this is not ablated.
  • Metric learning choices: The energy-weighted InfoNCE and cosine similarity are adopted without comparison to alternative metric-learning losses (e.g., triplet, N-pair, proxy-anchor) or learned Mahalanobis metrics; such comparisons could improve energy-distance fidelity.
  • Uncertainty-aware selection: The pipeline ranks by a point estimate (Dc) without uncertainty; ensemble JEPA models, Bayesian embeddings, or conformal selection to provide risk-aware candidate lists are not explored.
  • Diversity vs. quality trade-offs: Selection optimizes proximity to references (stability) but does not explicitly enforce chemical/structural diversity; integrating diversity-aware criteria (e.g., determinantal point processes or coverage constraints per chemical system) is an open avenue.
  • Sensitivity to hyperparameters and selection size: While k and N are varied, no principled, data-driven procedure is provided to choose them; adaptive policies (e.g., based on estimated precision/recall or uncertainty) could improve robustness.
  • Embedding space geometry: The appropriateness of Euclidean distance in the learned space is assumed; learning or validating a chemically informed metric (e.g., per-system scaling, local anisotropy) could reduce mis-rankings—this remains uninvestigated.
  • Canonicalization and representation choices: The SVD-based lattice parameterization can exhibit discontinuities under near-degenerate singular values; the effects of cell choice (primitive vs. conventional), atom ordering, and canonicalization on embedding stability are not analyzed.
  • Long-range interactions and periodic images: Crys-JEPA focuses on atoms within the unit cell without explicit modeling of inter-cell interactions; extensions that encode longer-range periodic effects (e.g., via neighbor lists or Ewald-inspired features) are open.
  • Data requirements and training costs: The paper reports inference-time advantages but not pretraining cost/efficiency for Crys-JEPA versus MLFFs; a thorough cost-benefit and data-scaling analysis (including forces/stresses supervision) is missing.
  • Novelty and uniqueness measurement robustness: Novelty/uniqueness rely on fingerprint distances that can be sensitive to cell choices and tolerances; more robust prototype-level deduplication and novelty definitions (e.g., via structure matcher or symmetry-aware metrics) should be evaluated.
  • Potential data leakage: JEPA pretraining on MP2022 may overlap with evaluation references (MP-20/Alex-MP-20). A strict train/eval split ensuring no leakage—and measuring its effect on both stability and novelty—has not been documented.
  • Property-conditional discovery: The framework focuses on stability; extending JEPA to multi-objective embeddings (e.g., stability + band gap/ionic conductivity) and demonstrating multi-objective screening/refinement remains an open challenge.
  • Threshold selection for AEA E: The fixed stability threshold (0.1 eV/atom) may not generalize across chemistries; methods to adapt thresholds per system or to target metastable regimes with quantified synthesis likelihood are needed.
  • Fairness across chemistries: The selection may favor chemistries abundant in the training data; explicit fairness or coverage constraints ensuring exploration of underrepresented chemical systems are not considered.
  • Evaluation beyond two benchmarks: Results are limited to MP-20 and Alex-MP-20; assessing performance on additional, more challenging benchmarks (e.g., larger cells, diverse prototypes, 2D materials) is necessary to establish generality.
View Paper Prompt  View All Prompts 

Practical Applications

Practical, real‑world applications of Crys‑JEPA and the screening‑and‑refinement pipeline

The paper introduces Crys‑JEPA, an energy‑aware embedding that approximates formation‑energy differences, and a screening‑and‑refinement loop that upgrades crystal generative models by selecting promising candidates without expensive DFT. Below are actionable applications grouped by deployment horizon.

Immediate Applications

These can be deployed with today’s tools (e.g., pymatgen/ASE, Materials Project datasets, existing MLFFs like MatterSim/CHGNet) and the methods described in the paper.

  • High‑throughput pre‑screening of stable candidates without DFT
    • Sector: materials R&D across energy (batteries, photovoltaics), catalysis, electronics, ceramics
    • Workflow/product: add a “Crys‑JEPA score” step before DFT in existing screening pipelines to rank 104–105 structures in minutes to hours; integrate as a pymatgen/ASE plugin or a CI step in computational materials workflows
    • Assumptions/dependencies: trained Crys‑JEPA covers target chemistries/structure sizes; final DFT/experiment still required for confirmation
  • Reference‑agnostic stability scoring at training time
    • Sector: academia/industry software for generative materials design
    • Workflow/product: use embedding distance to in‑distribution training crystals as a proxy for hull stability when the final reference set is unknown or evolving
    • Assumptions: training set contains representative stable phases; embedding preserves energy ordering across chemical systems
  • Active‑learning‑like refinement of crystal generators without DFT
    • Sector: software for generative design platforms (e.g., MatterGen/FlowMM/ADiT users)
    • Workflow/product: “screen–select–fine‑tune” loop using Crys‑JEPA distances to select top‑k% generations for continued training, improving V.S.U.N with minimal extra compute
    • Assumptions: initial generator produces valid structures; JEPA‑based ranking correlates with downstream stability
  • Rapid triage of hypothetical structure databases
    • Sector: academia/consortia managing large repositories (e.g., enumerated prototypes, polymorph libraries)
    • Workflow/product: rank legacy/hypothetical crystals to prioritize which entries get MLFF relaxation/DFT first
    • Assumptions: embeddings are computed directly from unit‑cell structures; out‑of‑distribution chemistries may need re‑training
  • Cost‑effective exploration of compositions and subsystems
    • Sector: energy materials (solid electrolytes, cathodes/anodes), thermoelectrics, superconductors
    • Workflow/product: use Crys‑JEPA to compare candidates to subsystem references (per the paper’s reference‑set heuristic) and prune unstable regions early
    • Assumptions: subsystem selection logic aligns with phase‑diagram practice; relaxation remains necessary prior to experiments
  • Improved validity and novelty balance in generative campaigns
    • Sector: corporate labs aiming for “stable‑but‑novel” materials for IP
    • Workflow/product: embed a JEPA‑based selector to steer generation away from trivial known neighborhoods without sacrificing stability; monitor V.S.U.N uplift
    • Assumptions: novelty metrics are computed against appropriate, up‑to‑date corpora; embeddings are not overly biased to the training distribution
  • Visualization and education on energy‑aware crystal manifolds
    • Sector: education, lab onboarding, internal model auditing
    • Workflow/product: UMAP/TSNE plots of Crys‑JEPA space to teach stability regions, diagnose model mode‑collapse, and guide diversity sampling
    • Assumptions: dimensionality reduction preserves neighborhood structure qualitatively
  • Lightweight API/service for stability‑aware ranking
    • Sector: software tools vendors, platform teams
    • Workflow/product: “Crys‑JEPA‑as‑a‑Service” (REST/gRPC) for embedding computation and candidate ranking; CLI integration with LIMS/ELNs to prioritize synthesis queues
    • Assumptions: privacy policies allow structural data sharing; latency/throughput sized correctly for queue volumes
  • Greener compute policies for materials screening
    • Sector: policy/operations (HPC centers, funding agencies)
    • Workflow/product: adopt surrogate‑first triage (JEPA → MLFF → DFT) as a recommended practice to reduce HPC hours and carbon footprint
    • Assumptions: performance tracking confirms that surrogate triage does not systematically discard viable candidates
  • Benchmarking and reproducibility improvements
    • Sector: academic benchmarking, model evaluation frameworks
    • Workflow/product: incorporate JEPA‑based distance metrics and V.S.U.N reporting into leaderboards to compare generators under a consistent, low‑cost stability proxy
    • Assumptions: community consensus on proxy metrics; release of trained JEPA weights for reproducibility

Long‑Term Applications

These require additional research, scaling to broader chemistries/sizes, or integration with experimental and autonomous systems.

  • Foundation‑scale energy‑aware crystal embeddings
    • Sector: software platforms, cross‑institution consortia
    • Product: large JEPA models trained on broader datasets (forces/stresses, >20‑atom cells, alloys) as a standard substrate for many downstream tasks (stability, property prediction, retrieval)
    • Dependencies: access to diverse, curated datasets; scalable training; evaluation standards across elements and pressure/temperature regimes
  • Multi‑objective surrogate screening (stability + properties)
    • Sector: energy (ionic conductivity, voltage), electronics (band gap, mobility), catalysis (adsorption, TOF)
    • Product: joint embeddings that align with multiple properties, enabling Pareto‑front selection before high‑fidelity simulations
    • Dependencies: labeled data for additional properties; careful loss design to preserve trade‑offs
  • Fully autonomous discovery loops with minimal DFT
    • Sector: autonomous labs/robotics
    • Workflow: generator → JEPA ranking → targeted MLFF/DFT for top candidates → robotic synthesis/characterization → feedback to generator and JEPA
    • Dependencies: robust domain adaptation from in silico to experiment; uncertainty quantification and fail‑safe policies
  • Real‑time experiment scheduling and queue optimization
    • Sector: high‑throughput experimental facilities
    • Product: on‑the‑fly JEPA scoring to prioritize instrument time for most promising samples (synthesis, diffraction, spectroscopy)
    • Dependencies: integration with LIMS; validated correlation between JEPA rankings and experimental success
  • Design copilots and CAD for materials engineers
    • Sector: software for R&D
    • Product: interactive tools that surface “nearby” low‑energy regions in embedding space, propose edits, and assess stability impact in real time
    • Dependencies: human‑in‑the‑loop UX; explainability layers mapping embedding moves to structural changes
  • Exploration strategies powered by embedding geometry
    • Sector: research on generative/decision‑making algorithms
    • Product: reinforcement learning or diffusion guidance using JEPA distances for exploration–exploitation balance, curriculum learning, or reward shaping
    • Dependencies: theoretical/empirical calibration of embedding distance to true energy differences across regimes
  • Standards for surrogate‑based screening in policy and regulation
    • Sector: standards bodies, funding/publishing policies
    • Product: guidelines for using surrogates (performance thresholds, validation protocols, audit trails) prior to DFT and experimental confirmation
    • Dependencies: community validation and acceptance; documented error bounds and failure modes
  • Cross‑domain extension to molecular/organic crystals and pharmaceuticals
    • Sector: pharma (polymorph screening), organic electronics
    • Product: re‑trained JEPA variants for organics to accelerate polymorph discovery and stability assessment
    • Dependencies: domain‑specific datasets (including conformational flexibility), new augmentations, different reference conventions
  • Digital twins of materials R&D programs
    • Sector: enterprise R&D
    • Product: simulation environments that use JEPA‑based surrogates to rapidly simulate discovery campaign outcomes and budget allocation
    • Dependencies: validated macro‑level correlations between surrogate decisions and program‑level KPIs
  • Embedding/model marketplaces and interoperability
    • Sector: ecosystem/interop
    • Product: exchange of pre‑trained JEPA models and adapters; standardized APIs so labs can adopt/benchmark quickly
    • Dependencies: licensing/ IP frameworks; versioning and provenance tracking

Each application’s feasibility hinges on key assumptions: that Crys‑JEPA embeddings remain predictive for the target chemical space and unit‑cell sizes; that MLFF or DFT relaxation is available for top candidates; that training data are representative; and that final decisions still undergo first‑principles or experimental validation.

View Paper Prompt  View All Prompts 

Glossary

  • Chemical system: The set of elemental species defining a material’s chemistry. "Its chemical system is denoted by T1T_1--T2T_2--⋯\cdots--TkT_k"
  • Composition--energy space: A space where each material is represented by its composition and energy for thermodynamic analysis. "in composition--energy space."
  • Convex hull: The thermodynamic lower envelope that defines the most stable phases for given compositions. "The convex hull is then defined as the lower convex envelope in composition--formation-energy space."
  • Crys-JEPA: A crystal-focused joint embedding predictive model used to learn an energy-aware latent space for screening. "we introduce Crys-JEPA, a joint embedding predictive architecture for crystals"
  • De novo generation (DNG): Generating entirely new crystal structures without templates. "This motivates the task of de novo generation (DNG)~\cite{cdvae}, which aims to discover entirely new crystal structures without relying on predefined templates."
  • Density functional theory (DFT): A first-principles quantum mechanical method for computing material energies and properties. "standard stability evaluation typically relies on density functional theory (DFT), which is prohibitively expensive at scale."
  • Energy above hull: The energy difference between a structure and the convex hull at the same composition; measures stability. "The energy above hull is defined as"
  • Energy-aware latent space: An embedding space structured so that distances reflect formation-energy differences. "a joint embedding predictive architecture for crystals that learns an energy-aware latent space preserving formation-energy differences."
  • Flow matching: A generative modeling framework that learns transport maps via differential equations. "particularly diffusion~\cite{ddpm} and flow matching~\cite{lipman2022flow}."
  • Formation energy: The energy of a compound relative to its constituent elements, indicating thermodynamic favorability. "The formation energy per atom of C\mathbf{C} is defined as"
  • Fractional coordinates: Atom positions expressed relative to the unit cell basis, taking values in [0,1). "the atomic fractional coordinates X∈[0,1)N×3\bm{X}\in[0,1)^{N\times 3}"
  • Haar-uniform distribution: The uniform distribution over rotations with respect to the Haar measure on a compact group. "is sampled from the Haar-uniform distribution"
  • InfoNCE: A contrastive learning objective that brings positive pairs together and pushes negatives apart. "with an InfoNCE objective"
  • Joint Embedding Predictive Architecture (JEPA): A framework that predicts target embeddings from context embeddings to learn invariant representations. "a joint embedding predictive architecture (JEPA)"
  • Lattice matrix: A 3×3 matrix whose rows (or columns) are lattice vectors defining a crystal’s periodic cell. "the lattice matrix L∈R3×3\bm{L}\in\mathbb{R}^{3\times 3}"
  • Lower convex envelope: The minimal convex function lying below a set of points; used here to define the convex hull of formation energies. "the lower convex envelope"
  • Machine-learning force fields (MLFFs): Data-driven interatomic potentials that approximate energies and forces for materials. "another option for stability screening is to use machine-learning force fields (MLFFs)"
  • Materials Project: A large public database of computed materials properties used as a reference set. "such as the Materials Project~\cite{MaterialsProject}"
  • Phase diagram: A representation of stable phases across compositions (and conditions), derived from convex-hull analysis. "we construct a phase diagram"
  • SiLU: An activation function (Sigmoid Linear Unit) defined as x·σ(x), used in neural networks. "denotes SiLU(⋅\cdot) activation"
  • Singular value decomposition: A matrix factorization used here to parameterize the lattice into rotation and symmetric components. "we factorize L\bm{L} through singular value decomposition"
  • Special orthogonal group (SO(3)): The group of 3D rotation matrices with determinant 1. "where U∈SO(3)\bm{U}\in\mathrm{SO}(3) (special orthogonal group) is sampled from the Haar-uniform distribution"
  • Thermodynamic stability: A condition indicating a structure lies on or near the convex hull (low energy above hull). "we regard a crystal as thermodynamically stable"
  • UMAP: A dimensionality-reduction technique used to visualize high-dimensional embeddings. "and then reduce the embedding dimension using UMAP~\cite{mcinnes2018umap-software}."
  • Unit cell: The fundamental repeating building block of a crystal’s periodic structure. "the periodic arrangement of its fundamental repeating unit, the unit cell"
  • V.S.U.N.: A composite evaluation of Validity, Stability, Uniqueness, and Novelty for generated crystals. "we collectively denote as V.S.U.N."
View Paper Prompt  View All Prompts 

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Generate Now

Collections

Sign up for free to add this paper to one or more collections.

Sign Up

Tweets

Sign up for free to view the 2 tweets with 78 likes about this paper.

Sign Up for Free