Single-modality model
Learns deeply from one data type, such as sequences, structures, images, or fields.
Layer 7 — Scientific AI · Chapter 2
Scientific Data, Representations, and Foundation Models
Scientific AI begins with representation. A model cannot learn directly from “a protein,” “a storm,” or “a material.” The object or process must be encoded as sequences, graphs, coordinates, fields, images, measurements, text, equations, or some combination of these forms.
The choice of representation determines which patterns are easy to learn, which symmetries are respected, how much data are required, and whether the model can generalize beyond familiar examples.
Major forms of scientific data
| Data form | Examples | Important structure |
|---|---|---|
| Sequences | DNA, RNA, proteins, time series | Order, motifs, long-range dependence, evolution |
| Graphs | Molecules, reaction networks, biological pathways | Nodes, bonds, topology, local and global interactions |
| Three-dimensional structures | Proteins, crystals, anatomical geometry | Distance, orientation, symmetry, shape |
| Images and volumes | Microscopy, radiology, telescopes, remote sensing | Spatial hierarchy, scale, resolution, acquisition conditions |
| Fields and meshes | Fluid flow, temperature, stress, electromagnetic fields | Continuity, conservation, boundaries, differential equations |
| Spectra and signals | Mass spectra, spectroscopy, seismic and acoustic signals | Frequency, peaks, mixtures, noise, calibration |
| Scientific text and equations | Papers, notebooks, patents, formulas | Claims, evidence, provenance, symbols, assumptions |
| Experimental records | Protocols, instrument settings, outcomes, failures | Conditions, units, batches, operators, uncertainty |
Representation should respect the science
Scientific objects often have symmetries and constraints. Rotating a molecule in space does not change its chemical identity. Relabeling equivalent atoms should not change a prediction. Conservation laws constrain physical trajectories. Dimensional units limit which quantities can be meaningfully combined.
Architectures that encode these properties can learn more efficiently and make fewer impossible predictions. This is called incorporating inductive bias: the model is designed so that some scientifically plausible relationships are easier to represent than arbitrary ones.
Scientific foundation models
A scientific foundation model is trained broadly enough to support multiple tasks within one or more scientific domains. It may learn from unlabeled sequences, structures, simulations, images, literature, or experimental databases, and then be adapted for prediction, generation, retrieval, or control.
Multimodal model
Connects different forms of evidence, such as literature, molecular structure, assay results, and microscopy.
Generative model
Creates candidate objects or trajectories that satisfy learned constraints.
Predictive model
Maps scientific inputs to properties, outcomes, or uncertainty estimates.
Simulation data and synthetic evidence
When experiments are scarce, simulations can generate large datasets. Simulated data are valuable because conditions and hidden variables are known. Yet a simulation is only an approximation of the world. A model trained entirely on synthetic evidence may reproduce the simulator’s assumptions and errors rather than reality.
Effective systems often combine three sources: established scientific laws, high-volume simulation, and lower-volume physical measurement. The model learns broad structure from simulation and is calibrated against real experiments.
The neglected value of metadata
A measurement without context can be misleading. Scientific datasets require units, instrument type, calibration, protocol, sample preparation, environmental conditions, laboratory, time, and uncertainty. Missing metadata can cause a model to treat differences in procedure as differences in nature.
Negative and failed results are equally important. If only successful experiments are published, the dataset gives a distorted view of what works. Recording failure conditions can reduce repeated mistakes and improve models of experimental feasibility.
Data quality problems
- Inconsistent units, naming conventions, and identifiers
- Duplicate records and hidden overlap between training and evaluation data
- Selection and publication bias
- Measurements produced by changing instruments or protocols
- Labels inferred from weak proxies rather than direct observation
- Missing uncertainty and confidence information
- Restrictions caused by privacy, intellectual property, or national security
Knowledge integration
Scientific understanding is distributed across structured databases, papers, patents, notebooks, code, instruments, and human expertise. Retrieval systems can connect these sources, but retrieval alone does not reconcile contradictions or determine which evidence is trustworthy. Scientific models need provenance: the ability to trace a claim or prediction back to the data, method, assumptions, and version that produced it.
Strategic opportunities
- Domain-specific data cleaning, normalization, and metadata reconstruction
- Secure scientific data exchanges and federated learning systems
- Multimodal scientific search with provenance and evidence comparison
- Tools for detecting leakage, duplication, bias, and protocol drift
- Scientific model evaluation datasets created through prospective experiments
- Data infrastructure that captures unsuccessful experiments and tacit process knowledge
In scientific AI, the dataset is not merely fuel. It is a partial, historically produced record of how humanity measured the world.