Gigafactories for Science

At Argonne National Laboratory, a liquid-handling robot missed a narrow well and picked up the microplate instead. Battery solvents softened other plates. Condensation distorted spectra. More than half of the roughly 150 chemicals needed for the campaign had to be ordered, often from small vendors. The researchers found that sourcing, rather than machine learning, set the final limit on throughput. Robertson et al., Digital Discovery 2026, describes the construction of an air-free self-driving laboratory for battery electrolytes. The paper reports labware incompatibility, condensation artifacts, robotic misalignment, manual reagent preparation, and sourcing as practical constraints.

The team wanted an autonomous laboratory for air-sensitive battery chemistry. Before a model could choose useful experiments, people had to keep oxygen out, invent seals a robot could handle, find plastics that survived the solvents, and decide what a distorted measurement meant. One misplaced plate was funny. An unrecorded change in plate material would not be.

Computation can widen a search space by millions while a laboratory settles one plate at a time. Formal mathematics can end in a proof checked by a small kernel, and software often has a test suite. Chemistry has no universal equivalent for a solvent lot, a scratched well, or six hours of missing temperature history.

The physical world has no kernel. Constellations of Borrowed Light makes that claim in a sentence, and The Discovery Engine stops at it: the engine ends before matter, law, and care. This is where the body begins. What follows is what the body costs.

The gap is already visible in materials research. GNoME predicted about 2.2 million structures stable relative to prior databases. A-Lab attempted fifty-seven inorganic targets in seventeen days and realized thirty-six. Scientists later judged four additional claimed successes inconclusive. Szymanski et al., Nature 2023, reports 36 of 57 targets over 17 days. Its 2025 Author Correction removed four inconclusive successes and clarified that the targets were new to the prediction platform, not necessarily new to science. Merchant et al., Nature 2023, reports the GNoME figures. These are computational predictions, not synthesized materials. In a separate bounded loop, OpenAI and Ginkgo Bioworks reported more than 36,000 cell-free protein-synthesis reactions across six rounds under human supervision and a 40 percent cost reduction against their baseline.

An automated laboratory that keeps its failures inside one organization creates a valuable private memory. Shared infrastructure could let the record of a failed plate at one site inform the first attempt at another. That matters more as models produce more candidates than any one facility can test.

Scientific progress moves through four queues: candidate ideas, physical execution, scientific review, and action authorization. More models shorten the first queue. More robots shorten the second only when robotic capacity is the constraint and instruments, materials, and trained operators are ready. Adding machines without maintenance, reviewers, and transfer support lengthens another line.

A gigafactory for science begins with two facilities that can run the same bounded method against common references, pay for independent review, and let the receiving site decide under its own rules. The unit scales when a third site can qualify the same work cell without private reconstruction. Its output is transferable physical learning.

Here, giga describes repeatable capacity, not the size of one building. A huge campus full of bespoke machines would remain a huge bespoke laboratory. Scale begins when another site can use the method and its failure record without losing the context that made either interpretable.

A data center scales through repeatable units of power, cooling, compute, and network. A science gigafactory needs an equivalent unit for physical learning. Its work cell combines a bounded method, compatible instruments, trained staff, reference materials, a review standard, and local permission to run.

The comparison breaks in the work itself. A battery plant repeats a known process to make a known object. A scientific facility runs a process to discover which object or process works. An apparently identical run may change with the solvent lot, operator motion, sample history, or instrument state. Useful scale preserves those differences instead of averaging them into a production number.

Standardization can also make a bad question efficient. A work cell may freeze a weak measurement, reward easy throughput, or starve work that does not fit its instruments. A gigafactory needs room to retire a method, challenge its reference, and admit that some questions should never enter the line.

Staff can pin the digital plan. Physical execution remains stochastic. A compiler can check declared identifiers, calibration status, and machine-readable preconditions before it produces a plan with allowed branches. It cannot find a dependency no one recorded or decide whether an omitted variable matters.

Software can record who reviewed, scheduled, stopped, or released a batch. It cannot grant those powers. They come from institutional appointment, facility rules, law, licensure, contract, and consent.

Physical outcomes need more than success and failure.

The execution record captures what the contract required, including instrument identity, calibration, lots, controls, interventions, deviations, and measurement references. It can still miss the causal variable. The next site needs enough context to investigate what the first site could not see.

The factory therefore funds metrology as core capacity. Common references help two facilities locate disagreement. Method transfer asks whether a result survives a new instrument, operator, environment, and scale. A successful bench result still faces process validation, engineering lots, change control, and release by a quality authority.

Before anyone funds a new hall, a public program can test the category inside two existing facilities. It can choose one air-sensitive materials method, give both sites common references and a bounded comparison standard, and pay an independent reviewer. Each site remains in charge of its own work.

The contract pays for unsuccessful work. A scientific negative, a justified stop, or a transfer failure can satisfy the milestone when it helps the other site avoid a bad assumption. Paying only for positive findings would turn the review queue into a sales target.

Public science already has useful ancestors. Department of Energy user facilities allocate scarce equipment through merit review, provide staff and safety systems, and charge full cost for proprietary use while supporting open publication for non-proprietary work. National Science Foundation Materials Innovation Platforms combine synthesis, characterization, modeling, and data inside user facilities. Their rules reserve substantial instrument time for external users and support access to technical expertise. See the DOE Office of Science user-facility definition and the current NSF Materials Innovation Platforms program. The NSF program requires at least half of supported instrument operating time for external users and applies full cost recovery to proprietary research. These are institutional precedents, not existing gigafactories.

That model broadens access without dissolving expertise. Outside researchers can inspect the work, submit bounded proposals, receive training, and use instruments under facility rules. Small laboratories need reference materials, staff help, and secure evidence access as much as they need a public webpage.

The hardware should become less bespoke over time. NIST researchers have proposed common interfaces and off-the-shelf modules in response to the time and cost of custom autonomous laboratories. A sample holder is not glamorous. Neither is a rack standard. Both matter when capacity must survive a vendor change. Joress et al., NIST and Matter 2026, proposes a composable autonomous-materials laboratory ecosystem built from community-driven standards and off-the-shelf modules. NIST presents this as a proposal, not a deployed standard.

The operating budget pays technicians, calibration engineers, data stewards, safety and quality staff, reviewers, and maintainers. Workers can stop a run and report an incident without losing milestone credit. Shift load, maintenance debt, safety events, and reviewer backlog sit beside instrument utilization. The freezer does not care that the milestone review is Thursday.

When evidence must remain protected, the contract names its owner, the reason, the expiry rule, and an independent custodian. Another authorized institution still needs enough access to challenge or use the result. A hash can prove that sealed bytes stayed fixed. It cannot make them useful.

The pilot measures time to an interpretable cross-site result, reviewer minutes, transfer failure, safety stops, downtime, and access by outside teams. It stops if the method travels only through heroic reconstruction, if reviewers cannot keep pace, or if safety and false-admission limits fail. Then the new hall does not get funded.

Meridian is an illustrative public-benefit materials gigafactory, not a forecast. Meridian, Sentinel, and their staff are composite scenarios. The scenes test which contracts, people, records, and authorities the proposed system would require. Real agencies and legal roles retain only their present powers. It opens with fewer instruments than its planners wanted and more calibration work than they budgeted. Mara Ortiz, a senior technician, takes one liquid handler offline after finding pressure drift that the vendor’s acceptance test missed. A temperature logger failed during a freezer shipment, so she quarantines the samples and marks every planned run that depended on them. The missing history stays missing.

The first corridor studies air-sensitive battery electrolytes. Participating scientists and user representatives chose the target. The safety staff cleared the materials. Each facility declared its method version, calibration status, reference material, and permission before receiving a pinned plate map with allowed branches.

At the partner site, a technician sees viscosity rise during mixing and uses an adjustment permitted by the method. Another plate fails its control. Staff classify it as a technical failure rather than evidence against the electrolyte. A third plate produces an unexpected decay curve that passes the measurement checks. The producer submits the anomaly and the required run record.

A measurement specialist confirms that the reader stayed within its qualified range. The appointed scientific reviewer asks whether the signal persists with the common reference and whether the plate chemistry could explain it. She can accept, defer, or reject a finding for the corridor’s research record. She orders the review queue, not the laboratory.

Meridian receives the packet and the review decision. Its quality staff see that their plate material and reader geometry differ. Meridian approves a bounded transfer study under its own scientific and safety rules. The operator puts it on Monday’s schedule for Mara. A valid difference would narrow where the method applies. A failed transfer would define the bridge the two sites still lack.

Within weeks, Meridian runs short on reviewer time before it runs short on instruments. Buying another robot would make the imbalance worse. The corridor funds another reviewer and returns mechanically incomplete packets before they reach that queue.

If an electrolyte survives the research corridor, a manufacturing partner starts a separate process-development path. Engineers test scale, mixing, contamination, and process windows. They may produce non-clinical engineering lots. A quality unit controls material disposition and release. The research finding supplies evidence. It grants no manufacturing permission.

Meridian’s charter separates corridor goals, scientific review, facility management, and quality authority. User representatives have power over access, data custody, and benefit sharing. Published appointments, rotating terms, recusals, and an independent appeal body make those limits enforceable.

Sentinel is a smaller public-health corridor. Before a signal arrives, its laboratories and manufacturing partner negotiate sample custody, data localization, benefit sharing, credit, refusal rights, and the evidence each jurisdiction permits to travel.

One evening, a wastewater feed rises above its local background. A partial sequence resembles a wildlife sample from another jurisdiction. The on-call scientific reviewer checks the assay’s false-positive history and handling record, then recommends more sampling. A regional laboratory director with authority under the local compact approves it. Biosafety staff clear diagnostic-primer work. The manufacturer reserves capacity, but its quality unit and public authorities retain every later release decision.

The second sampling band is negative. A handling review traces the first signal to possible local contamination. The public-health authority declines escalation, the manufacturer releases the capacity, and Sentinel stops primer work. Staff preserve the costs, batches, and evidence that ended the response. The program pays for that stand-down because the new evidence no longer justifies continued work.

Closed facilities can move faster. One executive chain may own the model, instruments, staff, and decision. Private laboratories already market proprietary science-factory systems and learning loops. Lila describes its AI Science Factories as proprietary instruments joined to an AI model and a proprietary learning loop. This supports the existence of the closed-stack strategy, not the company’s scientific performance claims. They can learn from failed routes that competitors never see. Their scientists are paid to advance the internal program, not to save another laboratory from the same mistake.

Different fields will use public, private, and hybrid institutions. User facilities already combine open publication with full-cost proprietary work. Other corridors may use fixed embargoes, regulator escrow, shared calibration standards, or public summaries linked to controlled evidence. A public summary can become transparency theater. An embargo can become permanent. A dominant facility can define standards around instruments smaller sites cannot afford.

Public infrastructure needs staff time, reference materials, secure review, and enough instrument capacity to make access usable. Cross-border corridors need reciprocal governance and access to resulting countermeasures. A network that extracts samples from one jurisdiction and sends authority elsewhere has reproduced an old institution with better software.

A foundation or public program can fund the first unit now with one method, two existing facilities, common references, paid review, and a receiving site with local authority. The contract pays for negatives, transfer failures, and justified stops. Construction money follows only after the method travels.

Passing that test earns more capacity. The partners can standardize a work cell, add instruments, qualify a third site, and fund a standing review service. Protein engineering, crop science, and public health would need different cells and different authorities.

On Monday morning, Mara opens the partner site’s failed-run packet before loading the same solvent. Her team inspects the calibration history, switches the plate material, and records why. Because the first site has no authority over her run, its packet has to give her enough evidence to choose before the robot starts.