A search space nobody can enumerate
Pick three elements from the first eighty or so that are not radioactive or noble gases, and there are ternary chemical systems; four elements give about 1.6 million quaternary systems, and each system contains many possible ratios and, for each ratio, many possible crystal structures. The experimentally known inorganic crystals number a few hundred thousand. Even counting generously, the space of compounds nobody has ever made is many orders of magnitude larger than the space anyone has looked at.
For thirty years the answer has been computation. Density functional theory made it possible to calculate the energy of a crystal that had never been synthesised, and from about 2010 open databases — the Materials Project, OQMD, AFLOW and NIST's JARVIS — published those energies for hundreds of thousands of structures. Machine learning entered as a way to go faster still: learn the DFT energy so that millions of candidates can be screened for the cost of thousands of DFT calculations, then hand the survivors to a laboratory, possibly a robotic one.
This lesson is a briefing on where that stands in late 2026. It is written from the materials scientist's side: Lumenary Chemistry's machine learning for molecules and reactions covers molecular representations and fitting interatomic potentials, and its self-driving laboratories covers Bayesian optimisation and closed-loop hardware. Here the questions are the ones a solid-state chemist asks: is the predicted compound stable, is it new, and did the laboratory really make it?
Stability is a hull, not a number
The first thing to get right is what "stable" means for a predicted crystal, because the headline numbers of every large-scale study are counts of stable materials.
Definition— Formation energy
The formation energy of a compound, , is its energy per atom minus the composition-weighted energies per atom of its elements in their reference states. For a binary , , in eV/atom. A negative value means the compound is lower in energy than the separated elements.
Negative is not enough. A compound can be lower in energy than its elements and still be higher in energy than a mixture of other compounds of the same overall composition, into which it would decompose. Plot every known phase of a chemical system as a point (composition, formation energy) and take the lower convex hull of those points: the hull is the lowest energy achievable at each composition by any combination of phases, which is the 0 K version of the common-tangent construction you met in Phase diagrams.
Definition— Energy above the hull
A phase's energy above the hull, , is its formation energy minus the hull's energy at the same composition. Phases with lie on the hull and are thermodynamically stable at 0 K (within the method used to compute the energies); phases with are unstable with respect to decomposition into the hull phases on either side, by that many eV/atom. A newly predicted phase that lies below the existing hull would become a new vertex and push other phases off it.
Two consequences matter for everything that follows. First, is a relative quantity: it depends on every other phase in the database, so adding one new stable compound can make thousands of previously "stable" predictions unstable. Second, the energies that decide stability are small. The median of the metastable materials people actually make is a few tens of meV/atom; the error of a good machine-learned formation-energy model is also a few tens of meV/atom. The block computes a hull for an invented A–B system, then asks how often a model with a 30 meV/atom error would get each stability call right.
import numpy as np
# A made-up binary system A-B. x is the fraction of B; energies are formation energies in eV/atom
# relative to the pure elements (which sit at 0 by definition).
phases = { # name: (x_B, formation energy in eV/atom)
"A": (0.00, 0.000),
"A3B": (0.25, -0.180),
"A2B": (1/3, -0.235),
"AB": (0.50, -0.300),
"AB2": (2/3, -0.215),
"AB3": (0.75, -0.150),
"B": (1.00, 0.000),
}
def lower_hull(points):
"""Lower convex hull of (x, E) points by Andrew's monotone chain; returns hull vertices left to right."""
pts = sorted(points)
hull = []
for p in pts:
while len(hull) >= 2:
(x1, y1), (x2, y2) = hull[-2], hull[-1]
# drop the middle point if it lies on or above the line from hull[-2] to p
if (x2 - x1) * (p[1] - y1) - (y2 - y1) * (p[0] - x1) <= 0:
hull.pop()
else:
break
hull.append(p)
return hull
def e_above_hull(x, e, hull):
xs, es = zip(*hull)
return e - np.interp(x, xs, es)
hull = lower_hull(list(phases.values()))
print("phase x_B E_f (eV/atom) E above hull (meV/atom)")
for name, (x, e) in phases.items():
print(f"{name:5s} {x:5.2f} {e:9.3f} {1000 * e_above_hull(x, e, hull):7.1f}")
# Now pretend the energies came from a model with a typical error of 30 meV/atom.
rng = np.random.default_rng(0)
truth = {n: e_above_hull(x, e, hull) < 1e-9 for n, (x, e) in phases.items()}
trials, agree = 2000, {n: 0 for n in phases}
for _ in range(trials):
noisy = {n: (x, e + (rng.normal(0, 0.030) if 0 < x < 1 else 0.0)) for n, (x, e) in phases.items()}
h = lower_hull(list(noisy.values()))
for n, (x, e) in noisy.items():
agree[n] += (e_above_hull(x, e, h) < 1e-9) == truth[n]
print("\nwith 30 meV/atom of model error, how often is each phase classified correctly?")
for n in phases:
if 0 < phases[n][0] < 1:
print(f" {n:4s} truly {'stable ' if truth[n] else 'unstable'}: {100 * agree[n] / trials:5.1f} %")In this system A3B, A2B, AB and AB2 are all on the hull and AB3 sits 11.2 meV/atom above it. With 30 meV/atom of random error, the deep, unambiguous AB is classified correctly 97.9 per cent of the time, but A3B only 52.2 per cent, AB2 57.2 and A2B 63.6 — barely better than a coin — and the unstable AB3 is called correctly 68.3 per cent of the time. Bartel and colleagues showed the real-data version of this in 2020: models whose formation energies looked excellent predicted stability poorly, because the errors that matter are those of differences between neighbouring phases, and composition-only models make correlated errors that do not cancel.
Universal interatomic potentials
The workhorse of 2026 is not a generative model but a universal machine-learned interatomic potential: one neural network that returns energies, forces and stresses for almost any combination of elements, trained on millions of DFT calculations. You met fitted potentials in Atomistic simulation, and Lumenary Chemistry's machine learning for molecules and reactions builds the locality and symmetry arguments behind them from scratch; what changed after 2022 is scale. M3GNet (2022) and CHGNet (2023) were trained on the Materials Project's relaxation trajectories across about 89 elements; MACE-MP-0 (first released at the end of 2023) showed that a single equivariant model could run stable dynamics on solids, liquids, interfaces and molecules; and in 2024–2025 groups at Microsoft (MatterSim), Meta (the OMat24 data set and then UMA, trained on about half a billion structures spanning molecules, materials and catalysts), Orbital Materials, Seoul National University (SevenNet) and others pushed the data sets to hundreds of millions of structures and released many of the weights.
What they are used for in discovery:
- Relaxing and pre-screening candidates. A generated or substituted structure is relaxed with the potential in seconds, and only candidates predicted near or below the hull go on to DFT. On Matbench Discovery the best universal potentials now identify stable crystals several times better than random selection would, which is why the 2025 paper behind it concluded they are good enough to act as cheap pre-filters.
- Properties that need many energy evaluations: phonons and thermal conductivity, elastic constants, ionic diffusion in candidate electrolytes by molecular dynamics, amorphous and disordered structures, surfaces and interfaces.
- Fine-tuning. A universal model fine-tuned on a few hundred DFT calculations of one chemical system usually beats a model trained from scratch on that system, the same pattern as pre-training in language models.
The way they are used looks like this. The code is shown, not run: it needs ASE, PyTorch and the model's weights.
# Not run here: needs ase, torch and a universal potential (here MACE-MP-0, via the mace-torch package).
from ase.io import read
from ase.filters import FrechetCellFilter
from ase.optimize import FIRE
from mace.calculators import mace_mp
atoms = read("candidate.cif") # a generated or substituted structure
atoms.calc = mace_mp(model="medium", default_dtype="float64")
FIRE(FrechetCellFilter(atoms), logfile=None).run(fmax=0.02, steps=500) # relax atoms and cell
e_per_atom = atoms.get_potential_energy() / len(atoms)
# Next: compare e_per_atom with the competing phases on the same footing (same model, or DFT
# references corrected to it) to get an estimated energy above the hull; send the best to DFT.The known failure modes are the ones any fitted model has, plus some specific to atoms. Universal potentials tend to make energy surfaces slightly too soft, so phonon frequencies and barriers come out low unless the model is fine-tuned; they inherit every systematic error of the DFT functional they were trained on, including the poor description of strongly correlated transition-metal oxides; and they are least reliable exactly where discovery wants them, in chemistries and structure types thin in the training data. An energy from a universal potential is a hypothesis about a DFT energy, which is itself a hypothesis about nature.
Generating crystals: GNoME and MatterGen
A filter needs candidates. The classical generators are still strong: substitution (take a known structure and swap chemically similar elements, guided by statistics of which substitutions occur in known compounds), prototype decoration (fill known structure types with new elements) and random structure searching (throw atoms into a box with sensible constraints and relax, as in Pickard and Needs's AIRSS). Generative models, from the crystal diffusion autoencoder of 2021 to diffusion models and language models trained on crystal files, add a learned prior over what crystals look like.
GNoME (Google DeepMind, Nature, November 2023) combined both with active learning. Graph networks predicted energies of candidates made by substitution and by random structure search; the most promising were computed with DFT; the results were added to the training set; and the loop repeated. Over successive rounds the models' precision at predicting stability rose sharply. The paper reported 2.2 million structures below the hull of the data then available, of which about 381,000 lay on an updated convex hull, and that 736 of its predicted stable structures had been realised independently by other laboratories; DeepMind contributed hundreds of thousands of the structures to the Materials Project. The scale was genuine and the active-learning result is important: a model that keeps getting the DFT answer back gets dramatically better at the specific question it is asked.
MatterGen (Microsoft Research, Nature, January 2025) is a diffusion model that generates crystals directly: it starts from noise in the atom types, the fractional coordinates and the lattice, and denoises all three together, respecting periodicity and symmetry. Its key feature is conditioning: adapter modules fine-tuned on labelled data steer generation towards a target chemistry, symmetry, band gap, magnetic density or bulk modulus. The paper reported that its structures were more often stable, unique and new than those of earlier generative models and much closer to their relaxed minima, and as proof of concept the team, with collaborators in China, synthesised one generated target, TaCr₂O₆, designed for a bulk modulus of 200 GPa, and measured a value within about 20 per cent of the target. The code and weights are openly available.
Autonomous laboratories and the argument over claims
The A-Lab at Lawrence Berkeley National Laboratory joined these predictions to robotic solid-state synthesis: it chose precursors and heating profiles with models trained on the literature, mixed and fired powders, and identified the products from powder X-ray diffraction with machine-learned phase analysis followed by automated Rietveld refinement. Its 2023 Nature paper, published the same day as GNoME, reported 41 new compounds made from 58 targets in 17 days. Lumenary Chemistry tells the story of the closed loop in self-driving laboratories; the materials questions are what came next.
The critique (2024). Solid-state chemists led by Leslie Schoop at Princeton and Robert Palgrave at University College London re-examined the published diffraction data. They argued that automated refinement had not been checked closely enough to support the identifications, and, more fundamentally, that most targets were predicted as ordered compounds, with each element on its own crystallographic site, while the products were more consistent with known compositionally disordered phases in which those elements share sites. A disordered solid solution of known structure type is not a new compound in the sense chemists use. In parallel, Anthony Cheetham and Ram Seshadri examined a sample of GNoME's predictions against the criteria a chemist would apply to a claimed discovery — is it new, is it credible, could it be useful — and reported finding little evidence of compounds that met the full set.
The correction (2026). In January 2026 the A-Lab authors published an author correction in Nature. They wrote that "new" had been meant as new to the prediction platform, not necessarily new to science, and changed the text accordingly; after a manual re-analysis of the diffraction patterns, peer-reviewed after publication, they confirmed 36 of 40 reported successes with 4 inconclusive and removed those 4 from the count; and they removed one compound that had mistakenly been in the training data.
MatterGen's example (2026). In a Materials Horizons communication, Mikkel Juelsholt argued that the synthesised TaCr₂O₆, which formed as a disordered rutile, is crystallographically the same as a disordered Ta–Cr–O rutile phase reported in the early 1970s and present in MatterGen's training data. The generative model, on this reading, predicted an ordered version of something already known, and the ordering is an artefact of how DFT represents disorder: it can only compute ordered cells.
Misconception
"The critiques show that AI materials discovery does not work." They show something narrower and more useful. The predictions, the robots and the diffraction were all real; what failed was the step that turns a measured pattern into a claim about a new compound, and the definition of "new" that did not account for disorder. Both are fixable: diffraction analysis with honest uncertainty and human review, structure matching that treats an ordered prediction and its disordered parent as the same material, and novelty checked against the experimental literature (the ICSD and the Pauling File), not only against a computed database.
Positive examples exist, and they share a pattern: a narrow, well-posed target and expert synthesis. In 2024 a team from Pacific Northwest National Laboratory and Microsoft screened about 32 million candidates with machine-learned models and physics-based filters for solid lithium-ion conductors, then synthesised and characterised a promising family, , as a solid electrolyte. The value of the computation was in narrowing a search for a property chemists already knew how to measure. Battery materials returns to solid electrolytes.
Measuring a screening model honestly
A model used as a filter should be scored as a filter. Suppose a campaign has candidates, of which a fraction (the prevalence) are truly stable. The model selects some; of those selected, the fraction truly stable is the precision; of all the truly stable, the fraction selected is the recall. The figure of merit Matbench Discovery uses is the discovery acceleration factor, : how many times more stable materials you find per DFT calculation than by picking candidates at random. Mean absolute error, the usual regression score, is close to irrelevant on its own, because what matters is the error near the decision threshold at .
The block simulates a campaign of 200,000 candidates with an invented but typical distribution of true , then screens it with models of increasing random error and, finally, with a model that is systematically biased.
import numpy as np
rng = np.random.default_rng(42)
n = 200_000
# True energies above the hull (eV/atom) of hypothetical candidates. The shape is made up but
# typical of a screening campaign: a few per cent at or below the hull, a long tail above it.
e_true = rng.gamma(shape=1.5, scale=0.12, size=n) - 0.015
stable = e_true <= 0.0
prevalence = stable.mean()
print(f"{n:,} candidates, {stable.sum():,} truly stable (prevalence {100 * prevalence:.1f} %)")
def screen(e_pred):
chosen = e_pred <= 0.0 # send everything predicted stable on to DFT
hits = (chosen & stable).sum()
return chosen.sum(), hits / chosen.sum(), hits / stable.sum()
print("\nmodel error DFT runs precision recall DAF")
for sigma in (0.02, 0.04, 0.08, 0.16):
runs, precision, recall = screen(e_true + rng.normal(0, sigma, n))
print(f"{1000 * sigma:5.0f} meV/at {runs:8,} {100 * precision:5.1f} % {100 * recall:5.1f} % {precision / prevalence:5.1f}")
noise = rng.normal(0, 0.04, n)
print("\n40 meV/atom of random error, with and without a systematic bias:")
for bias in (0.0, -0.03, +0.03):
runs, precision, recall = screen(e_true + noise + bias)
print(f" bias {1000 * bias:+4.0f} meV/at: {runs:6,} DFT runs, precision {100 * precision:4.1f} %, "
f"recall {100 * recall:4.1f} %, DAF {precision / prevalence:4.1f}")With a prevalence of 3.2 per cent, a 20 meV/atom model has a precision of 41.5 per cent and a DAF of 13.0; at 80 meV/atom the DAF falls to 3.9, and at 160 meV/atom to 2.3. Recall barely moves (from 62.5 to 51.3 per cent), because random error pushes about as many stable materials out of the selection as near-stable ones into it; what collapses is precision, which is what costs DFT time. The bias rows show a subtler point: a model that underestimates energies by 30 meV/atom recovers 82.4 per cent of the stable materials but sends twice as many candidates to DFT, while one that overestimates them is more precise and misses most of the discoveries. The threshold, and the direction of the bias, is a choice about cost, and a paper that reports only one operating point hides it.
Real benchmarks add three things this toy leaves out. The test set should be generated the way a real campaign generates candidates (Matbench Discovery uses structures from a substitution-based search, unseen in training); prediction should start from the unrelaxed structure, because a model that needs the DFT-relaxed geometry has already used the DFT it claims to save; and the hull must be built from the same database the campaign would use.
A property model on composition features
When the structure is not known — the usual situation at the start of a search — models work from composition alone. The classic recipe, from Ward and colleagues' Magpie features of 2016, is to describe a composition by statistics (mean, range, deviation, and so on) of elemental properties weighted by the atomic fractions, then fit a standard regressor. The block builds the smallest useful version for 24 tetrahedrally bonded semiconductors and insulators — group IV elements, III–V and II–VI compounds, plus MgO — whose room-temperature band gaps are well established, and predicts the gap from three features you can compute from the periodic table: the electronegativity difference, the mean period and the difference in valence electron count. It fits the logarithm of the gap, because gaps span a factor of 45 and relative errors are what matter.
import numpy as np
# Element data: Pauling electronegativity, period (row) and number of valence electrons.
ELEMENTS = {
"C": (2.55, 2, 4), "N": (3.04, 2, 5), "O": (3.44, 2, 6),
"Mg": (1.31, 3, 2), "Al": (1.61, 3, 3), "Si": (1.90, 3, 4), "P": (2.19, 3, 5), "S": (2.58, 3, 6),
"Zn": (1.65, 4, 2), "Ga": (1.81, 4, 3), "Ge": (2.01, 4, 4), "As": (2.18, 4, 5), "Se": (2.55, 4, 6),
"Cd": (1.69, 5, 2), "In": (1.78, 5, 3), "Sb": (2.05, 5, 5), "Te": (2.10, 5, 6),
}
# Room-temperature band gaps (eV) of elements and 1:1 compounds, rounded, as in standard tables.
DATA = [
("C", "C", 5.47), ("Si", "Si", 1.12), ("Ge", "Ge", 0.66), ("Si", "C", 2.36),
("Al", "N", 6.2), ("Ga", "N", 3.4), ("In", "N", 0.7),
("Al", "P", 2.45), ("Ga", "P", 2.26), ("In", "P", 1.34),
("Al", "As", 2.16), ("Ga", "As", 1.42), ("In", "As", 0.35),
("Al", "Sb", 1.6), ("Ga", "Sb", 0.73), ("In", "Sb", 0.17),
("Mg", "O", 7.8), ("Zn", "O", 3.37), ("Zn", "S", 3.6), ("Zn", "Se", 2.7), ("Zn", "Te", 2.26),
("Cd", "S", 2.42), ("Cd", "Se", 1.74), ("Cd", "Te", 1.5),
]
def features(a, b):
"""Features of the compound AB built from its two elements alone (no structure)."""
(xa, pa, va), (xb, pb, vb) = ELEMENTS[a], ELEMENTS[b]
return [1.0, # intercept
abs(xa - xb), # electronegativity difference: ionic character of the bond
(pa + pb) / 2, # mean period: bigger atoms, longer and weaker bonds
abs(va - vb)] # 0 for group IV, 2 for III-V, 4 for II-VI
X = np.array([features(a, b) for a, b, _ in DATA])
gap = np.array([g for _, _, g in DATA])
names = [a + b if a != b else a for a, b, _ in DATA]
def fit(X, gaps, lam=0.1):
"""Ridge regression on log(gap): gaps span a factor of 45, and errors scale with size."""
P = lam * np.eye(X.shape[1])
P[0, 0] = 0.0 # do not shrink the intercept
return np.linalg.solve(X.T @ X + P, X.T @ np.log(gaps))
def predict(X, w):
return np.exp(X @ w)
def rmse(pred, true):
return float(np.sqrt(np.mean((pred - true) ** 2)))
w = fit(X, gap)
print("log-gap weights: intercept %.2f, |dchi| %+.2f, mean period %+.2f, |dvalence| %+.2f" % tuple(w))
print(f"training RMSE {rmse(predict(X, w), gap):.2f} eV")
loo = np.array([predict(X[i], fit(np.delete(X, i, 0), np.delete(gap, i))) for i in range(len(gap))])
print(f"leave-one-out RMSE {rmse(loo, gap):.2f} eV, median error {np.median(np.abs(loo - gap)):.2f} eV")
base = np.array([np.delete(gap, i).mean() for i in range(len(gap))])
print(f"baseline (predict the mean of the others): RMSE {rmse(base, gap):.2f} eV")
print("\nhold out a whole family RMSE worst case")
families = {"nitrides": lambda a, b: b == "N", "oxides": lambda a, b: b == "O",
"antimonides": lambda a, b: b == "Sb", "group IV": lambda a, b: ELEMENTS[a][2] == 4}
for fam, member in families.items():
test = np.array([member(a, b) for a, b, _ in DATA])
p = predict(X[test], fit(X[~test], gap[~test]))
k = int(np.argmax(np.abs(p - gap[test])))
nm = [n for n, t in zip(names, test) if t][k]
print(f" {fam:12s} ({test.sum()}) {rmse(p, gap[test]):5.2f} eV {nm}: {p[k]:.2f} predicted, {gap[test][k]:.2f} measured")Three features and four weights take the error from 1.91 eV (always predicting the mean) to 0.95 eV in leave-one-out cross-validation, with a median error of 0.42 eV: the heavier the atoms, the smaller the gap (a weight of −1.28 on the mean period), and II–VI compounds have larger gaps than III–V compounds of the same row. That is roughly the accuracy a careful chemist gets from the same trends by eye, and it took no structures.
Then the honest test. Hold out every antimonide and the model, having seen arsenides and phosphides, predicts them to 0.29 eV: interpolation. Hold out every nitride and the error is 1.80 eV, with AlN predicted at 3.40 eV against a measured 6.2; hold out both oxides and ZnO comes out at 6.02 eV against 3.37. First-row anions are chemically special (small, very electronegative, short strong bonds), and a model that has not seen one cannot know it. A random split mixes nitrides into training and test alike and hides this; a grouped split that holds out a chemistry, a structure type or a later time period is the one that answers the question a discovery campaign asks.
Do not read physics into the weights. The log-gap weight on the electronegativity difference came out negative (−0.68), not because ionic bonds shrink gaps but because the valence-difference feature already carries most of the ionicity, and two correlated features share it in whatever proportion fits 24 points best. Feature importance in a correlated feature set is a statement about the data, not about the atoms. Real composition models use a hundred or more such features, tens of thousands of training compositions, and gradient-boosted trees or neural networks such as Roost that learn the weighting of elements themselves; their leave-one-out errors are smaller, and their grouped-split errors are still much larger than their random-split errors.
The professional view, late 2026
What practitioners in industry and national laboratories broadly agree on, as of late 2026:
- Universal potentials are infrastructure. They are used daily to relax candidates, run molecular dynamics of electrolytes and interfaces, compute phonons and pre-filter for DFT, and most new projects start by fine-tuning one rather than training from scratch. The open question is no longer whether they work but how to know when one is wrong: uncertainty estimates, active learning against DFT, and checks against experiment.
- Generation is cheap; verification is expensive. As in molecular design, any modern generator produces more plausible candidates than anyone can make. The bottleneck is credible stability and synthesisability prediction, and above all characterisation that can tell a new ordered compound from a known disordered one.
- Synthesisability is the real target. A compound on the hull may be unmakeable and a metastable one easy. Models trained on text-mined recipes are improving, and autonomous laboratories record failed syntheses routinely, which may prove their most valuable output.
- Claims need the norms of solid-state chemistry. Structure matching that handles disorder, novelty checked against the ICSD and the Pauling File as well as computed databases, diffraction refinements with their residuals and difference curves published, and independent replication. Since 2024 journals and authors have moved in this direction, and the debates above are part of why.
- Investment is large. Start-ups founded in 2024–2025 in the United States and Europe combine large models with autonomous laboratories, alongside programmes at AI laboratories, chemical companies and national laboratories in China, Japan, Korea and Europe. Judge them all by published structures, diffraction data and measured properties.
For your own work the toolkit is open: pymatgen for structures, phase diagrams and structure matching, ASE for driving calculators, the Materials Project, OQMD and JARVIS for data, Matbench Discovery for honest comparisons, and the open weights of most of the universal potentials named above. Statistical learning in general is Lumenary CS's learning from data, and how to evaluate models without fooling yourself is its evaluation.
Common misconceptions
Misconception
"381,000 new stable materials means 381,000 new materials." It means 381,000 structures computed to lie on a DFT convex hull, which is a strong and useful prediction but not a synthesis, a measured structure or a property. Some will be unmakeable, some are ordered versions of known disordered phases, and a few may prove to be the important ones.
Misconception
"A lower formation-energy error means a better discovery model." Stability depends on energy differences between neighbouring phases near the hull, and on errors that may be correlated across a chemical system. Judge a filter by precision, recall and discovery acceleration on a realistic, unseen candidate set, starting from unrelaxed structures.
Misconception
"If the diffraction pattern matches the predicted structure, the material has been made." A powder pattern constrains the lattice and the average arrangement; an ordered prediction and its disordered parent can give nearly the same pattern, and automated refinement can converge on a plausible but wrong answer. Identification needs refinement statistics, difference curves, sometimes other techniques (electron diffraction, spectroscopy), and a person who knows the chemistry.
Exercises
1. AI-assisted materials discovery
Warm-upHumanA screening campaign restricts itself to 30 elements. How many distinct quaternary chemical systems (sets of four different elements, ignoring ratios and structures) does it have to consider?
2. AI-assisted materials discovery
CoreHumanIn a binary system A–B, the only known compound is AB () with formation energy eV/atom. A model predicts a new compound AB₂ () with formation energy eV/atom. What is AB₂'s energy above the hull, in meV/atom, to the nearest integer?
3. AI-assisted materials discovery
CoreHumanIn a candidate set, 2.5 % of structures are truly stable. A machine-learned filter selects 4,000 candidates for DFT, of which 600 turn out to be stable. What is the filter's discovery acceleration factor (precision divided by prevalence)?
4. AI-assisted materials discovery
CoreHumanWhich of these were raised in the published critiques of, or the authors' 2026 correction to, the A-Lab's 2023 claims? Select all that apply.
5. AI-assisted materials discovery
CoreHumanYou trained a composition-based band-gap model on 5,000 oxides, sulfides, selenides and halides, and you want to know whether it can guide a search for new nitride semiconductors. Which validation answers that question?
6. AI-assisted materials discovery
CoreComputationalWrite energy_above_hull(phases) for a binary system A–B. phases maps a name to (x, E_f), where x is the fraction of B and E_f the formation energy in eV/atom. The pure elements at (0, 0) and (1, 0) are always part of the system even though they are not in the dict. Return a dict from each given name to its energy above the lower convex hull (0 for phases on the hull).
The tests use the lesson's system and a few small cases, including a compound with a positive formation energy and one that lies above the line joining two others.
7. AI-assisted materials discovery
StretchComputationalComposition-only models describe a material by statistics of elemental properties weighted by atomic fraction. Write composition_features(comp), where comp maps elements to amounts (integers or fractions), returning a tuple (mean, range, mad) of the Pauling electronegativity: the atomic-fraction-weighted mean, the range (largest minus smallest value among the elements present) and the fraction-weighted mean absolute deviation from the mean. The table ELECTRONEGATIVITY is provided. The same composition written with different multiples (Li₂O and Li₄O₂) must give identical features.
Further reading and resources
- Scaling deep learning for materials discovery · Merchant, Batzner, Schoenholz, Aykol, Cheon, Cubuk · 2023GNoME, Google DeepMind's graph networks trained by active learning; the source of the 2.2 million structures and 381,000 hull-stable entries.
- Millions of new materials discovered with deep learning · Google DeepMind · 2023The laboratory's own account of GNoME.
- A generative model for inorganic materials design · Zeni et al. (Microsoft Research) · 2025MatterGen, a diffusion model over atom types, coordinates and lattice, with property-conditioned fine-tuning.
- MatterGen source code
- An autonomous laboratory for the accelerated synthesis of inorganic materials · Szymanski et al. · 2023The A-Lab at Lawrence Berkeley National Laboratory.
- Author Correction: An autonomous laboratory for the accelerated synthesis of inorganic materials · Szymanski et al. · 2026Revises the novelty wording and the count of compounds made after a manual re-analysis of the diffraction data.
- Challenges in High-Throughput Inorganic Materials Prediction and Autonomous Synthesis · Leeman, Liu, Stiles, Lee, Bhatt, Schoop, Palgrave · 2024The detailed critique of the A-Lab's phase identification and of ordered-versus-disordered predictions.
- Artificial Intelligence Driving Materials Discovery? Perspective on the Article: Scaling Deep Learning for Materials Discovery · Anthony K. Cheetham, Ram Seshadri · 2024Two solid-state chemists ask which of GNoME's predictions meet the criteria for a discovery.
- Continued challenges in high-throughput materials predictions: MatterGen predicts compounds from the training dataset · Mikkel Juelsholt · 2026Argues that MatterGen's synthesised example is a known disordered rutile compound present in the training data.
- A critical examination of compound stability predictions from machine-learned formation energies · Bartel, Trewartha, Wang, Dunn, Jain, Ceder · 2020Why accurate formation energies do not guarantee accurate stability calls.
- A framework to evaluate machine learning crystal stability predictions · Riebesell et al. · 2025The paper behind Matbench Discovery.
- Matbench DiscoveryThe leaderboard for models used as stability pre-filters, with discovery acceleration factors.
- A general-purpose machine learning framework for predicting properties of inorganic materials · Ward, Agrawal, Choudhary, Wolverton · 2016The Magpie composition features, statistics of elemental properties.
- A universal graph deep learning interatomic potential for the periodic table · Chi Chen, Shyue Ping Ong · 2022M3GNet.
- CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling · Deng et al. · 2023
- A foundation model for atomistic materials chemistry · Batatia et al. · 2023MACE-MP-0; published in the Journal of Chemical Physics in 2025.
- UMA: A Family of Universal Models for Atoms · Wood et al. (Meta FAIR) · 2025
- Accelerating Computational Materials Discovery with Machine Learning and Cloud High-Performance Computing: from Large-Scale Screening to Experimental Validation · Chen et al. (Pacific Northwest National Laboratory and Microsoft) · 2024A screen of about 32 million candidates that ended in a synthesised sodium–lithium chloride electrolyte.
- Crystal Diffusion Variational Autoencoder for Periodic Material Generation · Xie, Fu, Ganea, Barzilay, Jaakkola · 2021
- The Materials ProjectOpen DFT data, phase diagrams and energies above the hull for well over a hundred thousand materials.
- JARVIS (NIST)NIST's open database and tools for DFT, force fields and machine learning.
- OQMDThe Open Quantum Materials Database, Northwestern University.
- pymatgenThe Python library behind the Materials Project, including its phase-diagram and structure-matching code.