Property prediction
Train models for electronic, mechanical, thermal, chemical, or processing-related properties.
Autoresearch for materials prediction and discovery.
Use your measured or simulated data to search model, representation, and feature spaces against the properties your program needs to predict or optimize.
Where Reverie fits
Model research across composition, structure, and properties.
Materials programs often combine sparse measurements, expensive simulations, and a large choice of representations. Reverie helps teams search those modeling decisions systematically.
Surrogate modeling
Search fast predictive models that approximate expensive simulations or experimental measurements.
Candidate prioritization
Rank compositions or structures for the next simulation, synthesis, or experimental campaign.
Turn a materials modeling objective into a scalable, GPU-backed research program.
Target property
Define the properties, performance metrics, and physical or practical constraints that matter.
Materials data
Connect experimental measurements, simulations, structures, compositions, or multimodal records.
Model space
Specify representations, architectures, descriptors, features, and training methods Reverie may explore.
Accelerated search
Run model training and evaluation on GPU compute while results guide the next research direction.
Candidate outputs
Receive trained predictors, comparative results, prioritized outputs, and a report of the research process.
Example project
Improve property prediction across a structural materials dataset.
Reverie can explore how representation and model choices interact while keeping evaluation fixed to the program’s actual use case.
- Objective
- Reduce held-out property prediction error while meeting uncertainty and inference requirements.
- Dataset
- Compositions, structures, measured properties, simulation outputs, process metadata, and grouped validation splits.
- Action space
- Descriptors, graph representations, neural architectures, surrogate methods, losses, pretraining, and hyperparameters.
- Evaluation
- Composition- or structure-aware holdouts, uncertainty quality, extrapolation behavior, robustness, and compute cost.
- Outputs
- Trained predictors, uncertainty estimates, prioritized candidates, evaluation artifacts, and a research report.
Explore broadly.
Evaluate consistently.
Reverie helps teams search more possible modeling approaches while keeping every candidate tied to the same objective, validation structure, and constraints.
Representation-flexible. Explore the descriptors and learned representations appropriate to your materials data.
Compute-ready. Use GPU-backed training for candidates that would be impractical to compare manually.
Reproducible. Preserve model artifacts, configurations, metrics, and research lineage for follow-up work.
Delivered to your team
Reusable models with the evidence behind them.
Trained predictors
Model artifacts, representations, preprocessing, configurations, checkpoints, and inference requirements.
Candidate analysis
Comparative accuracy, uncertainty behavior, extrapolation tests, tradeoffs, and prioritized downstream candidates.
Research report
A structured account of the explored model space, key findings, failed directions, and selected outputs.