The bottleneck moved to the bench
Machine learning for molecules and reactions ended on a pattern: a model proposes cheaply, and an experiment verifies expensively. A synthetic chemist runs perhaps a few to a few dozen reactions a day; a model can propose ten thousand candidates before lunch. The rate-limiting step of discovery is no longer having ideas but testing them.
Two developments attack that bottleneck. High-throughput experimentation (HTE) runs many small reactions in parallel: pharmaceutical groups have screened catalysts and conditions in 96-, 384- and 1536-well plates since the 2010s, at nanomole scale. A self-driving laboratory goes further and closes the loop, so that the results of one round of experiments choose the next without a person in between. Robots make and measure; an algorithm decides.
Strip away the robots and the question is old: given what I have measured so far, which experiment teaches me most about where the optimum is? George Box and K. B. Wilson asked it about chemical-plant yields in 1951; what is new is that the answer can be computed and acted on every few minutes, around the clock. This lesson builds that decision-maker from scratch — a Gaussian-process Bayesian optimiser, the workhorse of the field — runs it on a simulated reaction whose kinetics you know from Chemical kinetics, and then surveys what real self-driving laboratories have achieved as of 2026, what makes their results trustworthy, and what still needs people.
Anatomy of a closed loop
Every self-driving laboratory, whatever it makes, runs the same cycle.
Make. Liquid handlers — pipetting robots — dose stock solutions into well plates or vials, and are mature and reliable. Dispensing solids at the milligram scale is much harder and remains a common reason to call a human. Flow reactors pump reagents through heated tubing, so reaction time becomes residence time, set by a pump rate; they suit fast, exothermic and photochemical reactions. Inorganic materials need furnaces and powder handling. Some laboratories use fixed robot arms; others use mobile robots that carry samples between ordinary instruments shared with people (see Lumenary Robotics' manipulation and grasping).
Measure. The loop is only as good as its assay: liquid chromatography with UV and mass-spectrometric detection (LC–MS), benchtop NMR, UV–visible spectroscopy, powder X-ray diffraction for solids — the methods and calibration statistics of Analytical chemistry, run automatically.
Learn and decide. A model of the response (yield, selectivity, conductivity) is updated with every result, and an algorithm picks the next experiment — the heart of this lesson.
Definition— Self-driving laboratory
A self-driving laboratory (SDL; also called an autonomous laboratory or, for materials, a materials acceleration platform) is an experimental system in which automated hardware carries out experiments and a decision-making algorithm uses their results to choose subsequent experiments, in a closed loop, within a goal and constraints set by people. High-throughput experimentation runs many experiments in parallel but may be planned entirely in advance; automation replaces hands; autonomy replaces the decision about what to do next.
Autonomy is a spectrum — several authors have proposed graded scales borrowed from self-driving cars — and almost every system in 2026 sits in the middle: the algorithm chooses conditions within a space people defined, on hardware people maintain.
Bayesian optimisation
The task is to maximise an unknown function — yield as a function of temperature, time, concentrations, catalyst and solvent — when each evaluation costs hours and material, measurements are noisy, and there is no formula for or its gradient. Bayesian optimisation has two ingredients: a surrogate model of what we believe about everywhere, including how uncertain we are, and an acquisition function that turns the belief into a choice.
Definition— Gaussian process
A Gaussian process is a probability distribution over functions in which the values at any finite set of points are jointly normally distributed. It is specified by a mean function and a covariance function, or kernel, . The squared-exponential kernel
says that the response varies by about overall and that points closer than the length scale have similar responses.
Given noisy observations with , conditioning a multivariate normal distribution on them gives the posterior at a new point in closed form. With and ,
The mean interpolates the data, smoothed according to the noise; the variance is small near measured points and returns to far from them. The method is older than its machine-learning name. As kriging it was invented to estimate ore grades between boreholes by the South African mining engineer Danie Krige in the early 1950s and formalised by Georges Matheron in France; Jonas Mockus in Lithuania proposed using such models to decide where to search for an optimum in the 1970s; and Jones, Schonlau and Welch's 1998 paper made the combination standard.
The acquisition function balances exploitation (measure where the mean is high) against exploration (measure where the uncertainty is large, because the optimum might hide there). The commonest choice asks how much each candidate is expected to beat the best result so far.
Theorem— Expected improvement
Let the posterior at be normal with mean and standard deviation , let be the best value so far, and let be a small margin. The expected improvement
where and are the standard normal distribution and density functions.
Proof
Write with , so the improvement is when and zero otherwise. Then . Since , the last integral is , and .
∎
The first term rewards a high mean and the second rewards uncertainty; their sum is never negative. Other acquisition functions strike the balance differently: the upper confidence bound with a tunable , Thompson sampling (draw one plausible function from the posterior and measure at its maximum), and batch versions that fill a whole well plate at once.
A reaction to optimise
An honest test needs a problem whose answer you know and the optimiser does not. Take a reaction (wanted) that competes with a side reaction and is followed by an over-reaction , all first order. With , the yield of B after time is
and each rate constant follows the Arrhenius equation. Give the side reaction the lowest activation energy (35 kJ mol⁻¹, so it dominates at low temperature) and the over-reaction the highest (125 kJ mol⁻¹, so it speeds up most when hot), and the wanted reaction 70 kJ mol⁻¹ in between. The search space is 20–160 °C and 0.1–1000 min, with time on a logarithmic scale as chemists usually explore it.
import numpy as np
R = 8.314 # J mol^-1 K^-1
def k(k90, Ea, T):
"""Arrhenius rate constant (per minute), given its value at 90 degC."""
return k90 * np.exp(-Ea / R * (1 / (T + 273.15) - 1 / 363.15))
def true_yield(T, t):
"""Percentage yield of B for A -> B (wanted), A -> D and B -> C, at T degC after t min."""
k1 = k(0.15, 70e3, T) # A -> B
k3 = k(0.012, 35e3, T) # A -> D, a side reaction with a lower barrier
k2 = k(0.004, 125e3, T) # B -> C, over-reaction with a higher barrier
ka = k1 + k3
return 100 * k1 / (k2 - ka) * (np.exp(-ka * t) - np.exp(-k2 * t))
temps = [20, 55, 90, 125, 160]
times = [0.1, 1, 10, 100, 1000]
print("yield of B (%) " + "".join(f"{t:>8} min" for t in times))
for T in temps:
print(f"{T:>6} degC " + "".join(f"{max(0.0, true_yield(T, t)):12.1f}" for t in times))
T = np.linspace(20, 160, 701)
t = 10 ** np.linspace(-1, 3, 801)
Y = true_yield(T[:, None], t[None, :])
i, j = np.unravel_index(np.argmax(Y), Y.shape)
print(f"\nbest: {Y[i, j]:.1f}% at {T[i]:.1f} degC and {t[j]:.0f} min")
print(f"share of the search space within 3 points of the best: {100 * (Y > Y[i, j] - 3).mean():.1f}%")
for levels, factors in [(5, 2), (5, 4), (5, 6)]:
print(f"a {levels}-level grid in {factors} variables: {levels ** factors:,} experiments")The table shows the shape that makes reaction optimisation hard: a diagonal ridge. Hotter needs shorter, cooler needs longer, and off the ridge the yield collapses. The best conditions, 84.7 % at 83 °C for 39 min, sit in a region covering only 2.3 % of the search space.
The 25 experiments in that table are a chemist's natural five-by-five grid, and the best of them gives 74.9 % — almost ten points short of what is possible, because the ridge runs between the grid points. A finer grid would find it, but count the cost: five levels in six variables is 15,625 experiments, and real reactions have catalysts, ligands, solvents, bases and concentrations as well as temperature and time.
Changing one variable at a time does little better: it finds a point on the ridge but cannot see that the ridge rises in a direction that changes both variables together — the reason Box and Wilson invented response surfaces.
One step of the loop, in one dimension
Fix the time at 40 min and suppose four experiments have been run, at 30, 70, 120 and 150 °C, each measured with an assay whose standard deviation is 2 percentage points. The code fits a Gaussian process (length scale 0.15 of the temperature range, prior standard deviation 30 points), computes expected improvement over a grid of candidate temperatures, "runs" the best candidate, and repeats once.
import numpy as np
from math import erf, sqrt, pi
def k(k90, Ea, T):
return k90 * np.exp(-Ea / 8.314 * (1 / (T + 273.15) - 1 / 363.15))
def true_yield(T, t=40.0): # the simulated reaction, now at a fixed 40 min
k1, k3, k2 = k(0.15, 70e3, T), k(0.012, 35e3, T), k(0.004, 125e3, T)
ka = k1 + k3
return 100 * k1 / (k2 - ka) * (np.exp(-ka * t) - np.exp(-k2 * t))
ELL, SF, SN = 0.15, 30.0, 2.0 # length scale (in units of the range), prior sd, noise sd (% yield)
def kern(a, b):
return SF**2 * np.exp(-0.5 * ((a[:, None] - b[None, :]) / ELL) ** 2)
def posterior(u, y, us):
m = y.mean() # constant prior mean: the average so far
L = np.linalg.cholesky(kern(u, u) + SN**2 * np.eye(len(u)))
alpha = np.linalg.solve(L.T, np.linalg.solve(L, y - m))
Ks = kern(u, us)
v = np.linalg.solve(L, Ks)
return m + Ks.T @ alpha, np.sqrt(np.maximum(SF**2 - (v * v).sum(0), 1e-12))
def ei(mu, s, best, xi=0.5):
z = (mu - best - xi) / s
cdf = np.array([0.5 * (1 + erf(x / sqrt(2))) for x in z])
return np.maximum((mu - best - xi) * cdf + s * np.exp(-0.5 * z * z) / sqrt(2 * pi), 0.0)
rng = np.random.default_rng(7)
T_obs = np.array([30.0, 70.0, 120.0, 150.0])
measure = lambda T, noise: np.clip(true_yield(T) + noise, 0, 100) # an assay never reports < 0%
y_obs = measure(T_obs, rng.normal(0, SN, T_obs.size))
scale = lambda T: (T - 20) / 140 # degC -> unit interval
grid = np.linspace(20, 160, 141)
for step in (1, 2):
mu, s = posterior(scale(T_obs), y_obs, scale(grid))
best = posterior(scale(T_obs), y_obs, scale(T_obs))[0].max()
a = ei(mu, s, best)
print(f"step {step}: observed", ", ".join(f"{T:.0f} degC -> {y:.1f}%" for T, y in zip(T_obs, y_obs)))
print(" T/degC mean +/- 2 sd EI")
for T in (40, 60, 80, 90, 100, 110, 140):
i = np.argmin(abs(grid - T))
print(f" {T:6.0f} {mu[i]:6.1f} {2 * s[i]:9.1f} {a[i]:7.2f}")
T_next = grid[np.argmax(a)]
print(f" next experiment: {T_next:.0f} degC (true yield there {true_yield(T_next):.1f}%)\n")
T_obs = np.append(T_obs, T_next)
y_obs = np.append(y_obs, measure(T_next, rng.normal(0, SN)))Read the first table as the optimiser does. At 60 °C the mean is fairly high (61.8 %) but still below the best measurement so far, 72 %, and the band is moderate; at 90 °C the band is very wide (± 41 points) but the mean is only 49.4 %. Expected improvement peaks at 80 °C, where a mean of 66.4 % and a band of ± 26 points combine to give 2.74, against 1.24 at 60 °C and 1.41 at 90 °C. The measurement there, 83.1 %, shrinks the band nearby to ± 4 points, and the second step moves on to 83 °C, where the true yield at this time is 84.7 %. Two experiments, well chosen, did the work of a ten-point temperature scan.
Closing the loop
Now the full problem: temperature and log time, a budget of 20 experiments, the first five at random and the rest by expected improvement over a 41 × 41 grid of candidates. The recommendation is the measured condition with the highest posterior mean, not the highest single reading, which may be high only by luck. To judge fairly, the campaign is repeated with 20 random seeds and compared with 20 campaigns of random search on the same budget.
import numpy as np
from math import erf, sqrt, pi
def k(k90, Ea, T):
return k90 * np.exp(-Ea / 8.314 * (1 / (T + 273.15) - 1 / 363.15))
def true_yield(T, t):
k1, k3, k2 = k(0.15, 70e3, T), k(0.012, 35e3, T), k(0.004, 125e3, T)
ka = k1 + k3
return 100 * k1 / (k2 - ka) * (np.exp(-ka * t) - np.exp(-k2 * t))
def conditions(u): # unit square -> (degC, min); time on a log scale
return 20 + 140 * u[0], 10 ** (4 * u[1] - 1)
def run(u, rng): # "do the experiment": hidden truth plus assay noise (sd 2 points)
return float(np.clip(true_yield(*conditions(u)) + rng.normal(0, 2.0), 0, 100))
ELL, SF, SN = 0.15, 30.0, 2.0
def kern(A, B):
d2 = ((A[:, None, :] - B[None, :, :]) ** 2).sum(-1)
return SF**2 * np.exp(-0.5 * d2 / ELL**2)
def posterior(X, y, Xs):
m = y.mean()
L = np.linalg.cholesky(kern(X, X) + SN**2 * np.eye(len(X)))
alpha = np.linalg.solve(L.T, np.linalg.solve(L, y - m))
Ks = kern(X, Xs)
v = np.linalg.solve(L, Ks)
return m + Ks.T @ alpha, np.sqrt(np.maximum(SF**2 - (v * v).sum(0), 1e-12))
def ei(mu, s, best, xi=0.5):
z = (mu - best - xi) / s
cdf = 0.5 * (1 + np.array([erf(x / sqrt(2)) for x in z]))
return (mu - best - xi) * cdf + s * np.exp(-0.5 * z * z) / sqrt(2 * pi)
g = np.linspace(0, 1, 41)
CAND = np.array([(a, b) for a in g for b in g]) # 1681 candidate conditions
def bayes_opt(seed, budget=20, n_init=5, trace=False):
rng = np.random.default_rng(seed)
X = rng.random((n_init, 2))
y = np.array([run(x, rng) for x in X])
while len(X) < budget:
mu, s = posterior(X, y, CAND)
best = posterior(X, y, X)[0].max()
x = CAND[np.argmax(ei(mu, s, best))]
X, y = np.vstack([X, x]), np.append(y, run(x, rng))
if trace:
T, t = conditions(x)
print(f" run {len(X):2d}: {T:5.1f} degC {t:7.1f} min -> measured {y[-1]:5.1f}%"
f" model's best so far {posterior(X, y, X)[0].max():5.1f}%")
x_best = X[np.argmax(posterior(X, y, X)[0])] # recommend by the model, not one noisy reading
return true_yield(*conditions(x_best)), x_best
def random_search(seed, budget=20):
rng = np.random.default_rng(seed + 1000)
X = rng.random((budget, 2))
y = np.array([run(x, rng) for x in X])
return true_yield(*conditions(X[np.argmax(y)]))
print("one campaign (5 random starting experiments, then the loop):")
yb, xb = bayes_opt(1, trace=True)
T, t = conditions(xb)
print(f"recommended: {T:.1f} degC, {t:.0f} min, true yield {yb:.1f}% (best possible 84.7%)\n")
bo = np.array([bayes_opt(s)[0] for s in range(20)])
rs = np.array([random_search(s) for s in range(20)])
for name, r in (("Bayesian optimisation", bo), ("random search", rs)):
print(f"{name:22s} 20 campaigns x 20 runs: median {np.median(r):.1f}%, worst {r.min():.1f}%, "
f"within 3 points of the best in {(r > 81.7).sum()}/20")The trace shows both halves of the trade-off. Some runs explore corners a chemist would have skipped (20 °C for 0.1 min, 160 °C for 1000 min); from run 14 the loop climbs the low-temperature end of the ridge, 83.5 % at 65.5 °C and 126 min, and refines around it. The recommendation, 69 °C for 126 min, has a true yield of 82.8 %, within two points of the optimum at a different place on a nearly flat ridge. Across 20 campaigns Bayesian optimisation reached a median of 82.9 % and never worse than 79.3 %; random search reached a median of 79.2 % and a worst case of 70.2 %; the chemist's 25-point grid reached 74.9 %.
Real campaigns add complications that open-source tools handle:
- Categorical variables. Ligand, solvent and base are choices, not numbers; they are encoded as one-hot vectors or chemical descriptors, and the encoding matters as much as the algorithm.
- Batches. A well plate runs 96 experiments at once, so the algorithm must choose a jointly informative batch, not 96 copies of the best point.
- Several objectives. Yield against cost, or against the environmental metrics of Catalysis and green chemistry. Multi-objective methods return a Pareto front of conditions where no objective can improve without another worsening, and people choose along it.
- Constraints. Some conditions are infeasible (a solvent boils, a solid precipitates and blocks a line) and must be learned as well as avoided.
BoTorch and Ax provide these in Python, EDBO was written for reactions, and Summit and Olympus provide benchmarks for comparing optimisers fairly.
Self-driving laboratories around the world
The literature mixes genuine advances with enthusiastic claims. A fair survey, as of 2026:
Industrial high-throughput experimentation is the mature base. Merck & Co. (MSD) showed nanomole-scale screening in 1536-well plates in 2015, and Pfizer in 2018 a flow platform that screened thousands of reactions at nanomole scale and ran the best at micromole scale. Most large pharmaceutical companies now have HTE groups, but most HTE is still planned by chemists and executed by robots: high throughput, little autonomy.
Organic synthesis. Lee Cronin's group in Glasgow built the Chemputer (2019), a modular robot driven by χDL, a programming language for chemical procedures, and software that translates published procedures into it (2020). At MIT, Connor Coley, Klavs Jensen, Timothy Jamison and colleagues joined computer-aided synthesis planning to a reconfigurable flow platform (2019). Abigail Doyle's group at Princeton benchmarked Bayesian optimisation against expert chemists and released EDBO (2021). At the University of Liverpool, Andrew Cooper's group put a mobile robot into an ordinary laboratory: it ran 688 experiments in eight days in a ten-variable space and found photocatalyst mixtures for hydrogen production six times more active than its starting formulations (2020). Their 2024 system had mobile robots carry samples between a synthesis platform, an LC–MS and a benchtop NMR spectrometer shared with human researchers, and judged reactions on both measurements, as a chemist would.
Materials. Curtis Berlinguette's group at the University of British Columbia (2020) and a group in Tokyo (2020) built self-driving laboratories for thin films; the University of Science and Technology of China built a robotic "AI-Chemist" (2022) and used it to make oxygen-evolution catalysts from Martian meteorite material (2023). Lumenary Materials continues the story in AI-driven materials discovery.
Institutes and companies. Shared facilities include the Acceleration Consortium in Toronto, Liverpool's Materials Innovation Factory and programmes at United States national laboratories, with others in Europe and East Asia. Hardware ranges from instrument makers such as Chemspeed in Switzerland to lower-cost liquid handlers with an open-source Python control library; cloud laboratories such as Emerald Cloud Lab run experiments specified as code. Well-funded start-ups announced in 2025 aim at autonomous science at industrial scale; judge them, like everyone else, by published, reproducible results.
Language-model agents. Coscientist (Carnegie Mellon University, 2023) and ChemCrow (EPFL and the University of Rochester) showed that a language model can write the code that drives a liquid handler or a cloud laboratory (Machine learning for molecules and reactions; for agents in general, Lumenary CS's reasoning and agents). As of 2026 such agents help write, check and connect protocols, while the inner optimisation loop is usually a statistical method like the one above: cheaper, more predictable and easier to audit.
Example— When the loop's definition of success is wrong
In 2023 the A-Lab at Lawrence Berkeley National Laboratory reported that, in 17 days of continuous operation, it had synthesised 41 new inorganic compounds out of 58 targets predicted to be stable by large-scale calculations, planning recipes with models trained on the literature and judging success by automated analysis of powder X-ray diffraction patterns. In 2024, solid-state chemists led by Leslie Schoop (Princeton) and Robert Palgrave (University College London) published a detailed reanalysis. They argued that automated Rietveld refinement of the diffraction data was not yet reliable enough to support the identifications, and that the predicted compounds, all with each element on its own crystallographic site, had in about two thirds of cases most likely formed as known compositionally disordered versions, in which elements share sites — so that, on their reading, no new materials had been shown.
The lesson is not that autonomy failed; the robots made what they made. It is that a loop inherits its definition of success from its characterisation step and its model of what "new" means, and an error there is repeated at machine speed. The debate also showed science working: published data let independent experts check the claims within months.
Reproducibility and data standards
Automation is often sold as a cure for irreproducibility, since a robot doses the same volume every time. It cures some problems and hides others: a robot applies a systematic error consistently, and a loop that never records why it chose an experiment cannot be audited. Three practices make automated results trustworthy.
Design against hidden variables. Wells at the edge of a plate evaporate faster and sit at slightly different temperatures than wells in the middle. If each catalyst occupies one column, the catalysts in the outer columns share that bias, and the effect masquerades as chemistry. The cure is the one R. A. Fisher gave agricultural field trials at Rothamsted in the 1920s: randomise, and record the position.
import numpy as np
rng = np.random.default_rng(2026)
rows, cols, n_cat = 8, 12, 12
true = np.full(n_cat, 70.0)
true[6] = 74.0 # catalyst 7 really is the best, by 4 points
r, c = np.divmod(np.arange(rows * cols), cols)
edge = (r == 0) | (r == rows - 1) | (c == 0) | (c == cols - 1)
print(f"{edge.sum()} of {rows * cols} wells are on the edge of the plate")
def one_plate(layout):
if layout == "by column":
cat = c.copy() # catalyst k fills column k
else:
cat = rng.permutation(np.repeat(np.arange(n_cat), rows))
y = true[cat] + 5.0 * edge + rng.normal(0, 3.0, cat.size) # +5 points in edge wells
means = np.array([y[cat == k].mean() for k in range(n_cat)])
# with well positions recorded, fit yield = catalyst effect + edge offset by least squares
D = np.column_stack([cat[:, None] == np.arange(n_cat), edge]).astype(float)
fitted = np.linalg.lstsq(D, y, rcond=None)[0]
return np.argmax(means), np.argmax(fitted[:n_cat]), fitted[-1]
for layout in ("by column", "randomised"):
picks = [one_plate(layout) for _ in range(2000)]
naive = np.mean([p[0] == 6 for p in picks])
model = np.mean([p[1] == 6 for p in picks])
offset = np.mean([p[2] for p in picks])
print(f"{layout:10s}: best catalyst found by plain averages {100 * naive:5.1f}% of the time,"
f" by a model with an edge term {100 * model:5.1f}% (edge offset estimated {offset:.1f})")With the column layout and plain averages, the truly best catalyst is picked only 38.9 % of the time, because catalysts 1 and 12, entirely on the edge, come out almost level with it. Randomising the layout raises that to 87.8 %. And in either layout, a model that includes the well position recovers the right answer about 97 % of the time and measures the edge effect, 5.0 points — but only because the position was written down.
Record everything, including failures, in a standard form. The FAIR principles ask that data be findable, accessible, interoperable and reusable. For reactions, the Open Reaction Database schema captures reagents with amounts and sources, conditions, work-up, analyses and outcomes, including reactions that gave nothing, which journals almost never record and models badly need. Instruments have SiLA 2 for software control and the Allotrope Foundation's formats for analytical results; Germany's NFDI4Chem builds open notebooks and repositories such as Chemotion on such standards; and χDL records the procedure itself as executable code.
Make the decisions reproducible too. Record the optimiser's version, hyperparameters and random seed, and the model's state when each experiment was chosen. This lesson's code reruns to the same output because every random draw is seeded; a campaign should too.
What still needs people
A robot running all night tempts one to conclude that chemists are optional. The evidence of 2026 says otherwise:
- Choosing the question and the objective. An optimiser maximises exactly what it is told to. Told to maximise an HPLC peak-area ratio, it may find conditions that make a co-eluting impurity. The yield that matters — isolated, pure, reproducible at scale — is a different quantity, and judging which proxy is good enough is chemistry.
- Defining the search space. The loop explores the space it is given; the variables, ranges and candidate reagents are human choices, and important discoveries often come from changing the space rather than searching it.
- Safety. An autonomous platform can choose conditions no person would, so limits must be engineered in: maximum temperatures and pressures, incompatible-reagent rules, checks for heat and gas release, interlocks that stop a run, and a named person responsible. Hazardous chemistry is never left running without engineered controls and a risk assessment, for the reasons in Safety in chemistry.
- Keeping the hardware honest. Blocked lines, drifting calibrations, solids that will not dispense, reagents that decompose on the deck: much of an SDL group's daily work is maintenance and quality control, with the statistics of Analytical chemistry applied to every measurement the loop trusts.
- Interpreting surprises. An unexpected product or a result far outside the model's predictions is where discoveries hide, and an optimiser's instinct is to treat it as noise — one reason the Liverpool group's 2024 system judged reactions on several kinds of measurement rather than maximising one number.
There are questions for society too: who gets credit for an autonomously found result, how technicians' and chemists' jobs change, and how to make the tools available beyond the richest laboratories. Low-cost open-hardware SDLs, such as Baird and Sparks's "minimal working example" built around colour mixing and a light sensor, are one answer and good teaching tools.
Note
A safe loop to try at home. With an adult, pick a target colour and try to match it by mixing drops of three food colourings in cups of water, photographing each cup on a white plate under the same light. Record every mixture with its red, green and blue values from a colour-picker app, and choose each next mixture from what you have learned: a closed loop run by hand. Use only food colouring and water, don't drink the mixtures, and wipe up spills, since the dyes stain.
Common misconceptions
Misconception
"Self-driving laboratories replace chemists." They replace repetitive execution and some routine decisions. The objective, the search space, the safety limits, the hardware's reliability and the interpretation of anything surprising all remain human work, and the best-documented systems were built and supervised by teams of chemists, engineers and data scientists.
Misconception
"Bayesian optimisation always beats a good chemist or a grid." In two smooth variables with a generous budget, a grid or a chemist's intuition can do as well. Bayesian optimisation pays off when experiments are expensive, variables are many and interact, and consistency matters; with a badly chosen kernel or length scale it can do worse than random search. Benchmark before you believe.
Misconception
"The Gaussian process's error bars are the real uncertainty." They are the uncertainty under the model's assumptions: stationarity, smoothness at one length scale, Gaussian noise of a fixed size. Sharp cliffs (a precipitate forms), outliers (a blocked needle) and systematic errors (an edge effect) all violate them, and the posterior will be confidently wrong.
Misconception
"If the robot did it, it is reproducible." A robot repeats its own errors perfectly. Reproducibility comes from design (randomisation, controls, replicates), complete records in open formats, and independent checking — the A-Lab debate was settled by people re-reading the data, not by running the robot again.
Exercises
1. Self-driving laboratories
Warm-upHumanAn HTE group wants the full factorial screen of 3 palladium precatalysts × 4 ligands × 5 solvents × 2 bases, each at 4 temperatures. They use 96-well plates in a heating block that holds one plate at a time at one temperature. How many plates must they run?
2. Self-driving laboratories
CoreHumanThe best yield so far (by the model) is . A Gaussian process gives these posteriors for four candidate conditions. With , which has the largest expected improvement?
3. Self-driving laboratories
Warm-upHumanAn automated LC assay measures yield with a standard deviation of 3 percentage points per injection. How many independent replicate measurements must be averaged so that the standard error of the mean is at most 1 percentage point?
4. Self-driving laboratories
CoreComputationalWrite expected_improvement(mu, sigma, best, xi=0.0), which takes arrays of posterior means and standard deviations and returns the expected improvement of each candidate over best:
Where sigma is exactly 0 the formula divides by zero; there the improvement is certain, so return . Use math.erf for the normal distribution function: .
5. Self-driving laboratories
StretchComputationalWrite gp_posterior(x, y, xs, ell, sf, sn) for one-dimensional Gaussian-process regression with zero prior mean and the squared-exponential kernel . It returns two arrays, the posterior mean and the posterior standard deviation at the points xs:
The tests check, among other things, that one observation with large noise is shrunk towards the prior mean.
6. Self-driving laboratories
CoreHumanA self-driving laboratory finishes a 300-experiment campaign. Which of these should be in its permanent, machine-readable record for the campaign to be reproducible and reusable? Select all that apply.
7. Self-driving laboratories
CoreHumanIn 2023 an autonomous laboratory reported dozens of new inorganic compounds; in 2024 a group of solid-state chemists published a detailed reanalysis. What was the main substance of the critique?
Further reading and resources
- On the Experimental Attainment of Optimum Conditions · G. E. P. Box, K. B. Wilson · 1951Response-surface methodology, the sequential design of experiments that Bayesian optimisation generalises; it began with chemical-process yields.
- Efficient Global Optimization of Expensive Black-Box Functions · Donald R. Jones, Matthias Schonlau, William J. Welch · 1998The paper that made Gaussian-process surrogates and expected improvement standard.
- Nanomole-scale high-throughput chemistry for the synthesis of complex molecules · Buitrago Santanilla et al. · 2015Pharmaceutical high-throughput experimentation at Merck & Co. (MSD), with 1536-well plates.
- A platform for automated nanomole-scale reaction screening and micromole-scale synthesis in flow · Perera et al. · 2018Pfizer's flow-based screening platform.
- Controlling an organic synthesis robot with machine learning to search for new reactivity · Granda, Donina, Dragone, Long, Cronin · 2018
- Organic synthesis in a modular robotic system driven by a chemical programming language · Steiner et al. · 2019The Chemputer, University of Glasgow.
- A universal system for digitization and automatic execution of the chemical synthesis literature · Mehr, Craven, Leonov, Keenan, Cronin · 2020Turning written procedures into executable code (χDL).
- A robotic platform for flow synthesis of organic compounds informed by AI planning · Coley et al. · 2019
- A mobile robotic chemist · Burger et al. · 2020University of Liverpool. 688 experiments in eight days in a ten-variable space.
- Self-driving laboratory for accelerated discovery of thin-film materials · MacLeod et al. · 2020University of British Columbia.
- Autonomous materials synthesis by machine learning and robotics · Shimizu et al. · 2020
- Bayesian reaction optimization as a tool for chemical synthesis · Shields et al. · 2021Benchmarks Bayesian optimisation against expert chemists; the EDBO software.
- An all-round AI-Chemist with a scientific mind · Zhu et al. · 2022University of Science and Technology of China.
- An autonomous laboratory for the accelerated synthesis of inorganic materials · Szymanski et al. · 2023The A-Lab at Lawrence Berkeley National Laboratory.
- Challenges in High-Throughput Inorganic Materials Prediction and Autonomous Synthesis · Leeman, Liu, Stiles, Lee, Bhatt, Schoop, Palgrave · 2024A detailed critique of the A-Lab's claims; essential reading on characterisation in autonomous discovery.
- Autonomous mobile robots for exploratory synthetic chemistry · Dai et al. · 2024
- Autonomous chemical research with large language models · Boiko, MacKnight, Kline, Gomes · 2023
- The rise of self-driving labs in chemical and materials sciences · Milad Abolhasani, Eugenia Kumacheva · 2023
- Self-Driving Laboratories for Chemistry and Materials Science · Tom et al. · 2024A comprehensive review.
- What is a minimal working example for a self-driving laboratory? · Sterling G. Baird, Taylor D. Sparks · 2022
- The FAIR Guiding Principles for scientific data management and stewardship · Wilkinson et al. · 2016
- The Open Reaction Database · Kearnes et al. · 2021
- BoTorchOpen-source Bayesian optimisation in PyTorch; Ax (https://ax.dev/) is the experiment-management layer on top.
- EDBOExperimental design via Bayesian optimisation, written for reaction optimisation.
- SummitBenchmarks and strategies for reaction optimisation.
- SiLA 2An open standard for communicating with laboratory instruments.
- Allotrope FoundationData formats and models for analytical-instrument data.
- NFDI4ChemGermany's national research-data infrastructure for chemistry, including the Chemotion electronic laboratory notebook and repository.
- Acceleration Consortium, University of Toronto
- Materials Innovation Factory, University of Liverpool