Genesis Mission Portfolio III: Sixteen AI for Science Workflows at PNNL

First separate sixteen lead projects from twenty-six collaborations

PNNL says it leads sixteen Phase I projects in the first Genesis Mission selections and contributes to another twenty-six led by universities, companies or other national laboratories, for forty-two relationships in total. This brief maps only the sixteen PNNL-led projects. The other twenty-six help describe the collaboration network, but they are not additional PNNL-led projects and should not be counted as such.^[S3][S4][S2]

PNNL's role is broader than its project count

PNNL also co-leads the Genesis Mission's Models pillar and names staff responsible for model capabilities, baseline capabilities and commercial-cloud integration across the Transformational AI Models Consortium and American Science Cloud. These institutional roles help explain why its portfolio includes both domain science and operating-layer projects such as PORTUS and WASTECHAT. They show an intention to connect models, data, compute and facilities, not that a unified platform already exists.^[S5]

Placing sixteen projects back into the scientific workflow

An application-by-application list quickly becomes unwieldy. Reordering the projects by the role AI plays produces four groups: surrogates and agents for subsurface and Earth systems; design, manufacturing and materials loops; instruments, edge inference and operational monitoring; and knowledge and compute infrastructure. Each row also carries what the project is checked against (see section 04). The next-evidence column identifies what would help evaluate its target. It is not a completed or promised public deliverable.^[S4][S2]

Not one general AI scientist

PNNL's sixteen lead projects do not revolve around one general AI scientist. They insert different forms of AI into specific stages of research: neural operators stand in for costly simulations, agents orchestrate models and experiments, digital twins connect sensors to engineering decisions, edge models enter physical instruments, and knowledge systems organize decades of nuclear-waste records. PNNL's project descriptions are unusually specific about Phase I data sets, baselines and performance targets, which should make later verification easier. Those numbers are still targets, not completed results.^[S4]

A visible common thread is the use of physical constraints, established scientific codes, real sensors and expert review. In most proposals AI does not replace high-fidelity simulation or scientists. It screens scenarios, approximates expensive computation, proposes the next action, coordinates tools or escalates uncertain cases to more costly models and human judgment.^[S4]

Bottleneck

The roles of models, instruments, digital twins and research infrastructure were unclear.

New bottleneck

Reporting results against predefined hardware, field and held-out-data baselines.

What remains unsolved

Transfer across sites, cross-institution interfaces and real operating value.

Three provisional views

First, PNNL's portfolio looks more like a collection of research operating systems than a program focused only on larger models. Most proposals retain established scientific codes, experimental data and sensors, while AI compresses computation, connects tools or recommends the next action. That makes verifiers easier to define, but system integration, data quality and engineering reliability may become the main bottlenecks.

Second, the clearest near-term evidence may come from projects with hardware feedback or explicit physical baselines: quantum-amplifier tuning, LEGEND edge inference, fuel-cell manufacturing, molten-salt sensing and geothermal simulation. They can compare time, error, throughput, energy or anomaly detection directly. Cross-site transfer, scientific knowledge systems and distributed workflows will depend more heavily on longer-term use.

Third, the sixteen lead projects and twenty-six collaborations form an interface network, not forty-two independent accomplishments. The consequential question is whether workflows share data standards, model interfaces, compute and verification methods. If each remains a one-off vertical system, project count alone will not become platform capability.^[S3][S5]

Why rung 1

Rung 1, Exists: what is established is the selection itself (which teams, doing what, against which targets). On DOE and PNNL's public project pages, checked 2026-09-27, none of the sixteen projects shows a computed result, instrument data or field validation from this Phase I work, so none reaches rung 2. That says nothing against the teams' earlier published research; it is simply not a result of these projects. The baselines and numeric targets on the project pages make later checking easier, but a target is not evidence.^[S4][S1]

Selected for negotiation, not yet funded

PNNL calls these Phase I funding awards, while DOE's national notice also states that selection for award negotiations is not a commitment to issue funding and that negotiations may still be cancelled. The careful interpretation is that the teams have passed selection and entered the Phase I award process, while the public pages describe workflows to be developed, demonstrated and evaluated rather than completed scientific or engineering capabilities.^[S1][S3]

What the portfolio's composition signals

The workflow chart counts only the sixteen PNNL leads; the twenty-six collaborations are not in it. By topic, physical simulation and Earth systems has six, design, manufacturing and materials loops four, instruments, edge inference and monitoring four, and knowledge and compute infrastructure two. Topics show the spread, not which projects meet physical feedback: the watershed project is checked against measured streamflow and groundwater heads yet sits with simulation, while GOLD-FLOW sits with manufacturing although its page describes a CFD-driven loop with CFD or experimental checks on selected designs only.^[S2][S4]

So this brief gives each project a second tag for what it is checked against, one per project: the most physical check its page names. Hardware or a new experiment, seven: STEAM's new sensor-equipped tape-casting runs, synthesis and characterization of protein-vdW devices, the algae project's LEAPS pond simulator at month nine, LEAP-FWD's FPGAs, AQUA-WOLF's amplifiers at several institutions, SwiftFlow-EMT's hardware-testbed data and the molten-salt electrochemical measurements. High-fidelity simulation, four: geothermal, uranium recovery, EARL and GOLD-FLOW. Recorded field data, three: the E3SM hindcasts, the MSEEL fracture data and the measured watersheds. Software or offline tests, two: WASTECHAT and PORTUS.^[S4]

The edge cases, as the pages put them. The geothermal project borrows only the geometry of the Sanford Underground Research Laboratory test bed; its Phase I check is held-out simulated scenarios, with real-time field inversion left to Phase II. Uranium recovery builds its models from Lost Creek well logs, pump tests and historical pumping records, but its performance reference is PFLOTRAN simulation. SwiftFlow-EMT is mainly electromagnetic-transient simulation, supplemented by hardware-testbed data. The E3SM project is checked on the forecast skill of multi-year hindcasts against independent test data. The tags count plans, not results, and they are this site's reading, not PNNL's.^[S4]

Numeric targets improve testability, but they remain targets

PNNL's detailed pages provide several trackable metrics. The geothermal neural-operator project targets at least a 100-fold wall-clock speedup by month six and 1,000-fold by month nine. STEAM targets prediction error below 15 percent. The algae project aims to reduce productivity-deviation diagnosis from about twenty-five days to under twenty-four hours and lower seven-day forecast error by at least 30 percent. These numbers are valuable because they define success in advance. They must not yet be reported as achieved results.^[S4]

Later verification must look beyond a single accuracy number. Surrogate speedups need errors against high-fidelity baselines. Agentic workflows need explicit tool-calling and human-escalation rules. Instrument control must remain stable on real hardware. Digital twins need evidence that their recommendations improve field or manufacturing decisions. Speed without physical consistency, or an offline benchmark without an operating loop, is not enough.

PNNL's portfolio resembles research operating systems

Many projects retain established scientific codes, experimental data and sensors while AI compresses computation, connects tools or recommends the next action. The easy-to-miss risks are system integration, data quality and engineering reliability, not model size alone.

Specific targets make failure informative

Targets such as 100-fold speedup, 15% error or diagnosis within 24 hours are not yet results, but they define success in advance. If later results use the same baselines, missed targets will also reveal the real boundary of each approach.

What evidence would change the view

The first evidence layer is Phase I interim and final results: success and failure reported against predefined held-out data, hardware or field baselines, not only demonstration videos. The second is reuse: whether another team can run the software, data, interfaces and models on another instrument, site or resource configuration. The third is operating value: whether gains in speed and accuracy reduce expert time, experimental cost, facility downtime, manufacturing defects or decision risk.

Until those results appear, PNNL's sixteen projects are best treated as a relatively well-specified set of research hypotheses. They tell us where to look for the next evidence and make failure or missed targets informative too.