AI for Science Fundamentals

The common thread across every topic below: a scientific domain has an expensive, slow way to get ground truth (a wet-lab experiment, a physical simulation, a clinical trial), and ML offers a way to either replace a fraction of that expensive process with a fast prediction, or to search an astronomically large space of candidates (molecules, materials, protein sequences) for the few worth actually testing physically. That second pattern — a real ML filter narrowing an enormous candidate space before anything expensive happens — is genuinely computable, not just a description:

ML property-prediction threshold0.5

Candidate library200,000

Passed ML property filter31,805

Sent to physical synthesis/testing20

The exact same funnel shape applies whether the candidates are drug molecules, material compositions, or genomic variants -- the domain changes, this reduction pattern doesn't.

Real generated library of 200,000 candidate scores (a realistic long-tailed shape -- most candidates score low, a real minority score high). At ML threshold=0.50: 31,805 pass the filter (15.90% of the library) -- only the real top handful of those actually get expensive physical validation. This is the real, computed shape of "ML narrows an intractable search space before the expensive step," not just asserted.

Drug Discovery and Molecular ML

Finding a molecule that binds to a specific biological target (a protein implicated in a disease) traditionally means synthesizing and testing enormous numbers of candidate compounds physically — ML accelerates this at several stages: molecular property prediction (given a molecule's structure, predict its binding affinity, toxicity, or solubility without synthesizing it, using Graph ML since a molecule is naturally a graph — atoms as nodes, bonds as edges) dramatically narrows the candidate pool before any physical synthesis happens at all. This is squarely a Graph Classification-style task, just with molecules as the graphs and a biological/chemical property as the label.

Protein Structure Prediction

A protein's 3D shape (which determines its function) is determined by its amino acid sequence — but predicting that 3D structure from the sequence alone (the "protein folding problem") was a decades-old, extremely hard open problem, since a protein can fold into an astronomical number of possible shapes and experimentally determining structure (X-ray crystallography, cryo-EM) is slow and expensive. AlphaFold's breakthrough (built on attention mechanisms, see Attention & Transformers, applied to model relationships between amino acid positions) predicts structure directly from sequence at a level of accuracy competitive with physical experimental methods — one of the clearest, most consequential wins for deep learning in science to date, having genuinely accelerated structural biology research broadly, not just as a research demo.

Genomics

Applying ML to DNA/RNA sequence data — predicting gene expression from sequence, identifying disease-associated genetic variants, and increasingly, genomic foundation models (the same self-supervised pretraining approach as LLM Pretraining, applied to DNA sequences treated as a "language" with its own vocabulary and structure) that learn general-purpose sequence representations useful across many downstream genomic prediction tasks, the same pretrain-then-fine-tune pattern as NLP applied to a biological sequence instead of natural language.

Materials Science

The materials-science analog of drug discovery: searching for new materials with desired properties (superconductivity, strength-to-weight ratio, battery capacity) from an intractably large space of possible atomic compositions and structures. ML property-prediction models (structurally similar to molecular property prediction above, since a crystal structure is also naturally graph-representable) screen candidate materials computationally before the expensive step of actually synthesizing and testing them physically — directly analogous to drug discovery's screening funnel, applied to inorganic materials instead of organic molecules.

Physics-Informed Neural Networks (PINNs) and Neural Operators

PINNs: train a neural network to approximate the solution to a physical system's governing differential equations by incorporating the equations themselves directly into the training loss (penalizing the network's output for violating the known physics), rather than training purely on labeled input-output data — useful specifically when physical laws are well-known but data is scarce, letting the known physics constrain the network's predictions in regions with little or no training data. Real network, real ODE residual, one single labeled point:

Training objective

param a1

param b0

┈ true solution y=e⁻ˣ⬤ network prediction⬤ the one labeled point⬤ unlabeled physics-collocation points

Adjust a, b to minimize the real total loss (0.0000) by hand and watch how differently "good" looks with vs. without the physics term.

Real network output y(x) = e^(-ax) + bx·e^(-x), with only ONE labeled data point at x=0.1. Real data loss = 0.0000; real physics-residual loss (dy/dx + y, at 6 unlabeled points -- no ground truth needed, just the ODE itself) = 0.0000. With physics loss included, the fit is constrained everywhere, not just near the one labeled point.

Neural operators: rather than learning a single function (mapping one specific input to one specific output, as a standard neural network does), a neural operator learns a mapping between function spaces — e.g. learning to map any initial condition of a physical system directly to its solution function over time, generalizing across different initial conditions without needing to be retrained or re-solved from scratch for each one. Once trained, a neural operator can evaluate a new scenario orders of magnitude faster than re-running a traditional numerical simulation, at some accuracy cost — a genuinely different tool than PINNs (which approximate one specific solution) for the specific case of needing many fast approximate solutions across varying conditions. A real closed-form example of exactly this function-to-function shape — the 1D heat equation, solved for real, for any initial condition instantly:

Initial condition

time t0

┈ initial condition (t=0)⬤ real solution at time t

Real closed-form solution to the 1D heat equation (convolution with a real Gaussian kernel of width √(2·diffusivity·t)) -- the exact mapping a trained neural operator learns to approximate directly, function to function, without re-solving from scratch. Switch initial conditions and the SAME mapping (same "operator") applies instantly to the new input -- no retraining, unlike a standard network trained for one fixed input/output pair.

Simulation

Beyond neural operators specifically, ML broadly accelerates scientific simulation by learning fast, approximate surrogate models that stand in for expensive traditional numerical simulations (computational fluid dynamics, molecular dynamics, climate models) — trained on a set of expensive traditional simulation runs, then used to explore many more scenarios far faster than running the full traditional simulation for each one, trading some accuracy for a large speedup that makes previously-intractable large-scale scenario exploration feasible.

Molecular Generation and Docking

Molecular generation: rather than only screening existing candidate molecules (drug discovery, above), directly generate novel molecular structures with desired properties — reusing the same generative model machinery (VAEs, diffusion, and graph-generation-specific variants) covered elsewhere on this site, applied to molecular graphs as the object being generated instead of images or text.
Docking: predicting how (and how strongly) a candidate molecule physically binds to a target protein's 3D structure — a structure-prediction problem in its own right, increasingly tackled with learned models rather than purely classical physics-based simulation, directly consuming the protein structures AlphaFold-style models now make available at scale.

Scientific Foundation Models

The pretrain-then-fine-tune pattern from LLM Pretraining and NLP, applied within specific scientific domains: a large model pretrained (often self-supervised, in the same spirit as LLM pretraining) on a broad corpus of domain data — protein sequences, chemical structures, genomic sequences — producing general-purpose representations that get fine-tuned for many specific downstream scientific prediction tasks, rather than training a narrow model from scratch for each individual task. This mirrors exactly how a general-purpose LLM gets fine-tuned for many different downstream applications, just with the pretraining corpus and downstream tasks both drawn from a specific scientific domain instead of general text.

AI for Science section complete. Next: Domain AI Applications — where this section's biology/chemistry/physics focus sits alongside healthcare, finance, and other applied-AI domain tracks.