Researchers have unveiled four foundation models this week spanning electroencephalography, nuclear magnetic resonance spectroscopy, and 3D graphics — each demonstrating that pre-training on vast simulated or domain-specific datasets can produce systems that outperform narrower, task-specific predecessors and transfer knowledge across settings their creators never explicitly programmed.
NMR Spectroscopy: From Simulation to the Lab
A team led by Chen Yang and colleagues has introduced UltraNMR, a foundation model trained on 158 million paired simulated hydrogen and carbon NMR spectra. NMR spectroscopy is a core technique chemists use to deduce the structure of molecules, but experimental data is scarce — limiting previous AI tools to narrow, single-task applications.
UltraNMR tackles this by using simulated spectra for pre-training, then adapting to real experimental data. The researchers report state-of-the-art performance across a range of molecular analysis tasks and have built a retrieval library covering 94 million unique molecules. In a practical demonstration, the system helped identify the structures of two previously unknown natural products found in Chinese herbal medicines listed in the Chinese Pharmacopoeia — a result the team argues shows the model's real-world scientific utility.
EEG: Two Models, Two Approaches to Brain-Signal AI
Two separate research groups have tackled EEG foundation models — systems designed to learn generalisable representations of brainwave data and transfer that knowledge to clinical or brain-computer interface applications.
The B[FM]² project, from a team spanning MIT, KU Leuven, and other institutions, argues that existing EEG models fragment continuous brain rhythms by chopping signals into discrete patches or tokens. Their alternative trains directly on raw EEG waveforms using a technique called flow matching, without tokenisation or masking. Despite using roughly 30 times less pre-training data than competing models, B[FM]² claims top performance on seven of nine standard EEG classification benchmarks. Notably, synthetic EEG signals generated by the model were indistinguishable from real recordings by two board-certified neurologists.
A separate project, NeuroShield, focuses specifically on EEG-based identity authentication — using brainwave patterns as a biometric. The challenge it addresses is that EEG models typically break when moved between different headsets or recording setups. NeuroShield pre-trains a dual-stage transformer on data from 15,762 subjects across three public datasets, then fine-tunes for new devices. Tests on two previously unseen datasets show it reduces the equal error rate — a standard authentication accuracy measure — by between 0.44 and 8.06 percentage points compared to existing methods. The model has been released as open source.
3D Graphics: Teaching Machines to Think Like Artists
DreamUV addresses a less medically urgent but commercially significant problem: UV unwrapping, the process of flattening a 3D mesh's surface onto a 2D plane so textures can be applied. Professional artists develop strong stylistic preferences — straight seam lines, neatly aligned texture islands — that classical mathematical optimisation struggles to replicate.
The system, from researchers at Adobe Research and collaborators, frames UV unwrapping as a generative problem using flow matching, learning a distribution of artist-like layouts rather than a single solution. A user study with professional artists confirmed the outputs matched production standards. The researchers also introduce a training technique that accounts for errors introduced during the sampling process, improving stability.
A Common Thread
Across these four projects, a shared logic is visible: rather than engineering hand-crafted rules for specialist domains, researchers are using large-scale pre-training — often on simulated or synthetic data — to build flexible representations that adapt to real-world conditions. Whether that approach scales reliably across more demanding scientific challenges remains an active question in the research community.
Analysis
Why This Matters
- Foundation models are moving rapidly beyond language and images into specialist scientific domains — chemistry, neuroscience, and digital production — raising questions about how quickly AI tools could augment or change skilled professional work in these fields.
- The simulation-to-real strategy demonstrated by UltraNMR is particularly significant: it suggests that data scarcity in expensive experimental science may not be the bottleneck it once was, potentially accelerating drug discovery and materials research.
- Open-sourcing NeuroShield lowers the barrier for EEG-based authentication research, but also raises questions about the privacy and security implications of brainwave biometrics becoming more accessible.
Background
Foundation models — large neural networks pre-trained on broad data and then fine-tuned for specific tasks — became dominant in natural language processing after the release of GPT and BERT-style systems around 2018–2020. Their application to scientific domains followed: AlphaFold2 (2021) demonstrated the power of large-scale pre-training for protein structure prediction, inspiring analogous efforts across chemistry, genomics, and medical imaging.
NMR spectroscopy has long been considered one of chemistry's gold-standard tools for determining molecular structure, but its AI applications have lagged behind other spectroscopic methods partly because large curated experimental datasets are difficult and expensive to assemble. Simulated NMR data, generated from known molecular structures, offers a workaround — but bridging the gap between simulation and real instrument output has historically been problematic.
EEG-based AI has a longer history, with early clinical applications in seizure detection dating back decades. The push toward foundation models for EEG accelerated in the early 2020s as researchers recognised that task-specific models were being rebuilt repeatedly for each new dataset or device, wasting effort and limiting the field's progress. UV mapping in 3D graphics has similarly relied on decades-old geometric algorithms that work reliably but produce outputs lacking the polish of experienced human artists.
Key Perspectives
Researchers and developers: All four teams present their results as proof-of-concept advances with measurable performance improvements over prior systems. They argue that simulation pre-training and generalisation across devices are now tractable problems, not merely aspirational goals.
Clinical and scientific users: Practitioners in neurology and chemistry will likely welcome tools that reduce manual effort, but adoption in regulated environments — clinical EEG interpretation, pharmaceutical structure confirmation — will require extensive validation beyond benchmark performance. Real-world conditions are often noisier and more variable than test datasets suggest.
Critics and sceptics: Independent researchers have noted that benchmark performance on curated datasets does not always translate to real-world reliability. The claim that neurologists could not distinguish synthetic EEGs from real ones, for instance, raises questions about what that test specifically measured and whether it reflects clinically meaningful fidelity. Similarly, simulation-to-real gaps in NMR can be subtle and chemically significant.
What to Watch
- Whether UltraNMR's natural product identification results can be independently replicated by other chemistry laboratories using the model.
- Peer review outcomes for all four papers, which are currently available as arXiv preprints and have not yet undergone formal journal review.
- Regulatory and ethical frameworks for EEG biometric authentication as models like NeuroShield become more accessible — particularly around consent and the sensitivity of neurological data.