Pathways Forward · Field Survey

Nine Frontiers

Nine fields where AI systems moved in 2025–2026 from research curiosity to instrument of record — producing results other scientists now cite, build on, and occasionally have to correct.

Surveyed September 2026 Fields 9
Scope

AI-driven drug discovery — protein structure prediction, generative chemistry — is the part of this story that already reads as a headline, so it opens the log below. The eight fields after it are the other places the same underlying techniques — large-scale search, learned simulators, reinforcement learning against a physical or symbolic environment — are doing comparably real, if less-covered, work. Each field entry ends with a link to the full pathway curriculum for students who want to head toward it.

80–90%
Phase I success rate, AI-designed candidates vs. 40–65% historical

Computational Biology & AI Drug Discovery

Strand BIO

Isomorphic Labs, the DeepMind spinout built on AlphaFold 3's protein-structure predictions, advanced its first in-house oncology and immunology candidates toward first-in-human trials through 2025 and into 2026 — among the first therapies whose molecular design leaned on a deep-learning structure model rather than years of wet-lab screening alone. In February 2026 the company released IsoDDE, a follow-on model that roughly doubles AlphaFold 3's accuracy on the hardest ligand-binding cases, the ones where getting the geometry wrong wastes a trial rather than a simulation.

The early trial data is genuinely striking and genuinely partial: AI-designed candidates are clearing Phase I safety trials at 80–90%, well above the 40–65% historical average, but Phase II efficacy — does the drug actually work as well as hoped — is running at roughly the same rate for AI-designed and conventionally designed drugs alike. AI is proving it can design molecules that are safe; it hasn't yet proven it designs molecules that work better.

Caveat on record: as of September 2026, no AI-designed drug has received FDA approval. Every figure above describes trial performance, not a marketed medicine.
Pathway curriculum → Sources: IntuitionLabs, Philipp Dubach, AIM Media House
2.2M
candidate crystal structures, GNoME

Materials Science & Autonomous Discovery

Strand MAT

DeepMind's GNoME model screened candidate inorganic compounds and returned 2.2 million previously unknown crystal structures, roughly 380,000 of them stable enough to be plausible for real synthesis — at 80% precision, against roughly 50% for prior computational methods. Berkeley Lab paired the model with the A-Lab, a robotic synthesis line meant to close the loop from prediction to physical sample without a human choosing what to try next.

That loop is the field's actual frontier, and it isn't closed yet. In January 2026, Nature issued a correction to A-Lab's original claim of 41 autonomously synthesized novel compounds, after outside reviewers found several were already catalogued. The screening side of AI materials discovery is working; the synthesis side — getting a computer's stable-on-paper prediction into a lab dish with the predicted properties — is still the open problem.

Caveat on record: treat "autonomously discovered" materials claims from 2023–2025 as provisional until synthesis is independently verified — this field corrects itself in public, which is a feature, not a scandal.
Pathway curriculum → Sources: MIT Technology Review, Next Wave Insight
1,000×
less compute than physics-based forecasting, ECMWF AIFS

Climate & Atmospheric Modeling

Strand CLI

ECMWF's AIFS became the first AI system to enter operational forecasting service at a major weather center in February 2025, with NOAA following in January 2026. Google DeepMind's GraphCast, the model that opened this door, beats ECMWF's own deterministic system on 90% of benchmark targets, produces a ten-day global forecast in under a minute, and improves tropical-cyclone track accuracy by up to 20% — while running on roughly a thousandth of the energy a physics-based run consumes.

Nobody has retired the physics engines. Every major forecasting service now runs AI and numerical weather prediction side by side and blends them, because the learned models are fast and cheap but still occasionally miss extreme, out-of-distribution events the physics-based models were built to handle. The 2026 story here is operational maturity, not replacement.

Pathway curriculum → Sources: NOAA, Communications of the ACM
<1 ms
plasma-control decision latency, PACMAN / DIII‑D

Fusion Energy & Plasma Physics

Strand FUS

DeepMind and EPFL's Swiss Plasma Center first showed in 2022 that a reinforcement-learning controller could hold and reshape superheated plasma inside a tokamak in real time, including exotic configurations classical controllers weren't tuned for. The 2026 development is speed and foresight: the PACMAN framework, demonstrated on the DOE's DIII-D reactor, took control of heating systems and predicted edge energy bursts before they happened, and a related deep-RL controller now predicts and heads off tearing-mode instabilities — the disruptions that can end a plasma shot — before they start, where conventional controllers could only react after the fact.

DeepMind has since partnered directly with Commonwealth Fusion Systems, aiming this control work at SPARC, the tokamak CFS is building toward net energy gain. Fusion's hardest problems are still materials and confinement, but plasma control — keeping a 100-million-degree ring of gas in one place — is now substantially an AI problem.

Pathway curriculum → Sources: Princeton Plasma Physics Laboratory, U.S. Department of Energy, Google DeepMind
48
scalar multiplications, AlphaEvolve's 4×4 matrix algorithm

Mathematics & Automated Discovery

Strand PRF

AlphaEvolve, Google DeepMind's evolutionary code-search system, was set loose on more than 50 open problems across analysis, combinatorics, geometry, and number theory and improved on the previous best known solution in about a fifth of them. Its single most-cited result: a new algorithm for multiplying two 4×4 matrices in 48 scalar multiplications, breaking a record Strassen's algorithm had held since 1969.

The more consequential change is the pipeline it sits in. AlphaEvolve searches for a construction, Deep Think derives a proof of why it works, and AlphaProof formalizes that proof in the Lean theorem prover — a division of labor demonstrated end-to-end on the finite-field Kakeya problem. AlphaProof separately reached silver-medal-standard performance on International Mathematical Olympiad problems under competition rules. Mathematics is turning out to be one of the more tractable domains for AI, precisely because a proof is checkable.

Pathway curriculum → Sources: Nature, Google Research
weeks → hrs
chip floorplanning time, AlphaChip vs. human layout

Chip Design & Electronic Design Automation

Strand EDA

AlphaChip treats chip floorplanning — placing billions of components and the wiring between them — as a game a reinforcement-learning agent can play, and it has been used on the physical layout of Google's own TPU chips through several generations, cutting layout time from weeks of expert work to a few hours. What was a research curiosity in 2020 is now shipping product: by 2026, both dominant EDA vendors have folded RL-based placement and routing into their flagship tools — Cadence's Cerebrus and Synopsys's DSO.ai — and Nvidia uses its own internal ML tools on successive GPU generations.

What AI has not yet done is remove the human from the loop entirely: no autonomous agent has taken an industrial chip from specification to tape-out without engineers checking its work at a zero-tolerance error bar, where a single flaw can void an entire wafer run.

Pathway curriculum → Sources: Google DeepMind, EE Times
Sys 1 / Sys 2
dual-architecture reflex + reasoning, GR00T N1

Robotics & Embodied AI

Strand ROB

NVIDIA's Isaac GR00T N1 is the first open, generally licensed foundation model for humanoid robots — one model, fine-tunable across different robot bodies, rather than a bespoke policy per machine. Its dual-system design splits the job the way psychologists split human cognition: a fast "System 1" turns a plan into continuous motor commands at reflex speed, while a slower vision-language "System 2" reasons about the scene and what the instruction actually requires. NVIDIA trains it mostly in simulation on its Jetson Thor hardware, because collecting real-world robot trial-and-error data is slow and physically destructive at the volumes deep learning needs.

The test of whether this generalizes past demo reels is arriving on a schedule: Unitree is set to ship the first open reference humanoid built on GR00T and Jetson Thor to researchers in late 2026, which will be the first point outside NVIDIA's own labs that the model's claims get stress-tested at scale.

Pathway curriculum → Sources: NVIDIA Newsroom, arXiv
70–80%
herbicide reduction, John Deere See & Spray field trials

Agriculture & Precision Farming

Strand AGR

John Deere's See & Spray systems use onboard computer vision to tell crop from weed row by row and fire herbicide only at the weed, cutting chemical use 70–80% against blanket spraying in field trials — a rare case in this survey where the AI win is also a straightforward cost and input-reduction win a farmer feels immediately. Deere's 2025 acquisition of GUSS Automation, whose autonomous orchard and vineyard sprayers already operate across 2.6 million acres guided by GPS and LiDAR, extends the same approach to tree and vine crops that don't fit a row-crop tractor's geometry.

Full-cycle autonomous tractors — seeding, cultivating, harvesting, unsupervised — are Deere's longer-run bet, sold increasingly as a subscription on top of the hardware rather than a one-time purchase. Precision agriculture is the field in this survey where AI adoption is least visible and most already-arrived: it looks like an ordinary tractor from the road.

Pathway curriculum → Sources: Farmonaut, TechBuzz
741
DFT-confirmed superconductor candidates, of 1.3M screened

Energy Storage & Superconductor Discovery

Strand ENR

A complete AI-accelerated workflow for superconductor discovery, published in npj Computational Materials in 2026, trained a model called BEE-NET to predict a candidate compound's superconducting critical temperature directly from its structure — accurate to within roughly 1.6°F of the true value — and used it to cut 1.3 million candidate structures down to 741 that are both stable and confirmed by first-principles calculation to superconduct above −451°F, the cryogenic threshold the search was screening for. Separately, machine-learning structure search has already turned up working new superconductors in the lab, including YRu₃B₂ and LuRu₃B₂, a kagome-lattice pair identified computationally before being synthesized and tested.

The same screen-then-verify pattern is reshaping battery research: AI models now do the first pass on cathode, anode, and solid-state electrolyte candidates — predicting stability, ionic conductivity, and voltage before anyone runs a cell test — on chemistries like sodium- and graphene-based solid-state batteries. It's the same core move as the materials-science and drug-discovery entries above — a learned model narrows an impossibly large search space to a short list worth a chemist's bench time — applied to the specific problem of moving and storing energy without loss.

Pathway curriculum → Sources: npj Computational Materials, PMC, AI Discovers New Superconductors (2026)
Reading this

Two things repeat across all nine fields. First, the actual technique is rarely novel by 2026 — large-scale search, learned simulators standing in for expensive physics, reinforcement learning against a scoreable environment — only the domain is new; a student fluent in one of these fields' methods can move to another faster than the field-specific vocabulary suggests. Second, every field above still keeps a human or a classical system in the loop somewhere — a synthesis chemist, a numerical-weather-prediction ensemble, a verification engineer, a Lean proof checker. Where a field claims to have removed that check entirely is usually where to look hardest before taking the claim at face value.

Each pathway curriculum linked above is a full grades‑9-through-career sequence built off the same core — math, systems engineering, and machine learning — as the Construction Track this survey supports.