PROJECT SUMMARY

The blunt question behind this project is simple: when a scientific model is called a “foundation model,” does it actually transfer to new physics and new designs, or has it learned the training generator unusually well?

I am building a small open benchmark and release package around two research lines I already work on as a computational science PhD student at Seoul National University. The first is PDE-PFN, a prior-data fitted neural solver that uses in-context inference for partial differential equations. The second is a foundation-model approach for robust, generalizable metamaterial design. They look like different applications, but both face the same scientific test: performance on equations, coefficients, boundary conditions, and geometry families that were not represented in the easy part of the training distribution.

The output will be a practical package that a graduate student or small lab can inspect and run: held-out benchmark definitions, data-generation utilities, reproducible configurations, compute accounting, representative checkpoints where licenses permit, and a short report of both successes and failure cases. Any public release will respect existing coauthor, institutional, data, and software-license obligations.

GOALS AND EXECUTION

During the first two months I will freeze data manifests and held-out tests before the main scaling runs. For PDE-PFN this means varying PDE families, coefficients, spatial resolution, and boundary or initial conditions, including noisy approximate trajectories. For metamaterials it means separating structurally different geometry families and testing forward prediction as well as conditional inverse design.

Months three and four are for the expensive part: model, representation, robustness, and compute-efficiency sweeps. I will report not only best-case accuracy but also transfer gaps, invalid designs, achieved-property error, diversity, and the amount of compute used.

Months five and six are for clean reproduction, documentation, and release. The goal is not to promise a universal solver. It is to leave behind evidence that makes it harder to confuse interpolation with transfer, plus a reusable starting point for groups that cannot afford to rebuild the entire stack.

USE OF FUNDS

The minimum viable project is USD 5,000. That would cover a focused set of GPU experiments, storage, and release tooling for one carefully bounded benchmark.

At USD 15,000, the budget is approximately USD 10,000 for GPU compute, USD 2,000 for data generation and storage, USD 2,000 for research engineering and reproducibility work, and USD 1,000 for documentation and dissemination.

At the USD 25,000 maximum, I would expand the held-out families and independent reproduction runs: approximately USD 14,000 compute, USD 4,000 data/storage/validation, USD 4,000 engineering and release tooling, and USD 3,000 documentation, reproduction, and dissemination. Funds are for the project rather than tuition or general living expenses.

TEAM AND TRACK RECORD

I, Dongseok Lee, will lead and administer this project. I am a PhD student in the Interdisciplinary Program in Computational Science at Seoul National University. My recent work includes PDE-PFN, physics-informed neural PDE methods, data-efficient generative models for digital materials, and “Toward a robust and generalizable metamaterial foundation model.” Existing papers naturally have coauthors, but this funding proposal does not depend on recruiting a new mandatory collaborator, receiving a professor’s signature, or securing an institutional nomination.

FAILURE MODES

The main risk is that broad transfer does not appear: a larger model may still overfit the data generator. That is a useful result if the held-out tests and negative evidence are released clearly. A second risk is that compute limits reduce the number of ablations; the tiered budget makes the smallest version complete rather than merely unfinished. A third risk is release friction from licenses, data terms, or coauthor review. I will design the new benchmark code and manifests for release from the start, and where a specific artifact cannot be redistributed, publish the protocol, configuration, and aggregate evidence that can legally be shared.

FUNDING TO DATE

For this specific open benchmark package, I have not received dedicated cash funding. Compute-credit applications submitted for parts of the work are still pending and are not counted as funds raised here.