Explore the research

Energy

Can AI improve storage, grid operation and energy research, with real-world benefits?

17%

field summary estimate

6 outcomes · Assessed November 7, 2026 · Low confidence

The field estimate averages its outcome scores. It is a reading aid, not a measured fraction of the field that AI manages.

How the estimate built up.

Each band is one outcome's share of the field estimate, so the top line is the estimate itself. Every field starts at 0% on December 31, 2022. Dates before the latest are retrospective estimates. Point at a band to see that outcome's score on each date.

Storage performance35%
Research productivity30%
Operational efficiency0%
Net energy0%
Grid reliability0%
Fusion control35%

View exact historical values and reconstruction notes

What we measure

Each percentage is a judgment about progress toward independent advancement. Open the evidence to inspect the basis for it.

Fusion control

35%estimated progress

What we track

Plasma tracking error, stable operating duration, disruption rate, and experimental reproducibility on specified hardware.

Why this estimate?

Real-machine control experiments and fast feedback demonstrate useful bounded autonomy. Reproducibility across operating regimes and autonomous research planning remain limited.

Advancements

DeepMind and EPFL control TCV tokamak plasma with deep reinforcement learningFebruary 16, 2022. One neural network controlled all 19 magnetic coils of the TCV tokamak and produced several plasma shapes, including a 'droplet' configuration holding two plasmas in the vessel at once. Published in Nature. Predates the window; included as the landmark basis for learned plasma control.
DeepMind reports up to 65% better plasma shape accuracy for RL tokamak control21 July 2023. The arXiv paper 'Towards practical reinforcement learning for tokamak magnetic control' reports shape accuracy improved by up to 65%, a substantial reduction in long-term plasma-current bias, and training time for new tasks cut by a factor of 3 or more. The figures are simulation results, validated with new experiments on the TCV tokamak.
AI controller on DIII-D forecasts tearing instabilities up to 300 ms aheadFebruary 21, 2024. A Princeton-led team reported in Nature that a model trained only on past experimental data forecast tearing-mode instabilities up to 300 milliseconds in advance at the DIII-D tokamak, and a reinforcement-learning controller adjusted plasma shape and beam power in real time to avoid them.
ML-assisted 3D-field control gives burst-free high performance on KSTAR and DIII-DMay 11 (2024). A Princeton-led team reported in Nature Communications the highest fusion performance without harmful edge energy bursts on two tokamaks, KSTAR and DIII-D. A machine-learning surrogate model cut the needed computation from tens of seconds to milliseconds, allowing real-time adjustment of the magnetic field.
PACMAN makes fusion control decisions in milliseconds on DIII-DSeptember 2, 2026. PPPL reports PACMAN making control decisions in milliseconds and testing the framework in five DIII-D experiments.

Read the dated findings and sources

Status: Real-machine control demonstration. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

September 2, 2026: PPPL reports PACMAN making control decisions in milliseconds and testing the framework in five DIII-D experiments. PPPL: PACMAN fusion control experiments.

Interpretation and tracking value

A demonstration count and decision latency do not establish disruption rate, stable duration or tracking error across operating conditions.

Next comparable measurement

Record shot-level tracking error, stable seconds, disruptions/all shots and repeatability against the existing controller under matched plasma conditions. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Grid reliability

0%estimated progress

What we track

Unserved energy, outage duration, constraint violations, and renewable curtailment compared with established control methods.

Why this estimate?

The baseline cites a funded project rather than an achieved reliability result. No post-anchor autonomous progress toward this particular outcome can be scored from that evidence.

Advancements

National Grid reports 55% fewer outage minutes on circuits using AiDash tree analyticsNovember 6, 2024. National Grid in Massachusetts reported 30% fewer customers interrupted and 55% fewer outage minutes on circuits managed with AiDash's satellite-imagery AI vegetation platform, against 8-11% fewer interruptions from traditional cycle-based trimming, and about $2 million saved per year. The comparison is each circuit's three-year average versus the year after work; the article notes a small sample size.
Open Climate Fix says its AI solar forecast halved errors in Britain's grid control room7 November 2025. The National Energy System Operator (NESO) control room uses Open Climate Fix's Quartz Solar forecast to decide how much reserve to buy. Open Climate Fix says forecasting errors were halved since adoption, equivalent to at least GBP 30 million per year in imbalance costs. The figures are the developer's own and are not independently verified in the article.
DOE announces an $11.5 million AI electricity projectSeptember 1, 2026. DOE announced an $11.5 million Genesis-related electricity project in September 2026.

Read the dated findings and sources

Status: Development activity; measured field impact unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

September 1, 2026: DOE announced an $11.5 million Genesis-related electricity project in September 2026. DOE Office of Electricity: Genesis project.

Interpretation and tracking value

A project award is not an achieved reduction in outages, unserved energy or curtailment. No qualifying reliability effect is established by the announcement.

Next comparable measurement

Obtain operational trials with matched demand/weather; measure unserved MWh, outage minutes, constraint violations and renewable curtailment. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Net energy

0%estimated progress

What we track

Report plasma gain, facility energy balance, and net electricity delivered separately. Count all relevant energy inputs and auxiliary systems.

Why this estimate?

The cited plasma-control demonstration does not establish net energy or delivered electricity. Adjacent control progress earns no net-energy outcome credit.

Advancements

LLNL deep-learning model gave 74% ignition probability before the Dec 2022 NIF shotAug. 21, 2025. Lawrence Livermore reported in Science that a physics-informed deep learning model, used days before the December 5, 2022 experiment, predicted a 74% probability of ignition (greater than 70% chance of exceeding energy breakeven); the shot achieved ignition and its result fell within the predicted yield range. This is a prediction of target gain, not facility energy balance or delivered electricity.
PACMAN fusion-control system is demonstrated on DIII-DSeptember 2, 2026. PACMAN was demonstrated as a fusion-control system on DIII-D.

Read the dated findings and sources

Status: Control progress; facility net-energy result not established. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

September 2, 2026: PACMAN was demonstrated as a fusion-control system on DIII-D. PPPL: PACMAN fusion control experiments.

Interpretation and tracking value

Plasma-control performance is not plasma energy gain, facility break-even or net electricity. The cited demonstration supplies no qualifying net-electricity baseline.

Next comparable measurement

Meter plasma inputs/output, facility auxiliaries and delivered electricity separately; report duration and uncertainty without treating any one gain ratio as another. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Operational efficiency

0%estimated progress

What we track

Delivered kWh per unit input, operating cost per MWh, and losses under matched demand, weather, equipment, and reliability conditions.

Why this estimate?

The cited cooling improvement predates 2022 and therefore belongs to the requested starting baseline. No verified post-anchor increment is documented in this outcome file.

Advancements

DeepMind AI raises data-centre cooling savings to about 30%August 17, 2018. Google/DeepMind reported autonomous cooling savings improving from 12% to about 30% over nine months.
DeepMind and Trane RL control cuts cooling energy about 9% and 13% at two sites11 November 2022. The paper 'Controlling Commercial Cooling Systems Using Reinforcement Learning' reports live experiments at two real-world facilities with Trane Technologies, with energy savings of approximately 9% and 13% respectively. Submitted three weeks before the window opens (revised December 14, 2022).
Meta RL pilot cuts data-centre supply fan energy 20% and water use 4%September 10, 2024. Meta reported that a simulator-based reinforcement learning model controlling the supply airflow setpoint at one data-centre region reduced supply fan energy consumption by 20% and water usage by 4% on average across various weather conditions. The pilot began in 2021.

Read the dated findings and sources

Status: Measured historical operator-reported saving. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

August 17, 2018: Google/DeepMind reported autonomous cooling savings improving from 12% to about 30% over nine months. DeepMind: autonomous data-centre cooling.

Interpretation and tracking value

This is cooling-system performance at the operator's facilities, not a 30% reduction in total data-centre electricity or evidence of current fleet-wide savings.

Next comparable measurement

Compare weather/load-adjusted cooling and total-facility energy; include AI compute and maintenance, reliability and cost per delivered service. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Research productivity

30%estimated progress

What we track

Experiments and elapsed time per independently verified advance, with total human and compute effort.

Why this estimate?

Data-efficient battery lifetime prediction supports faster experimental selection, but external validation and full campaign resource accounting remain human-dependent.

Advancements

DeepMind's GNoME predicts 2.2 million crystals, including 528 lithium-ion conductorsNovember 29, 2023. GNoME predicted 2.2 million new inorganic crystals, about 380,000 of them stable, raising known stable materials from about 48,000 to 421,000. Candidates include 528 potential lithium-ion conductors for batteries. External researchers had independently synthesized 736 of the predicted materials, and Berkeley Lab's robotic A-Lab made more than 41 new materials.
Microsoft and PNNL screen 32.6 million candidates and build a new solid electrolyteJanuary 9, 2024. Microsoft's AI and cloud computing narrowed 32.6 million candidate materials to about 500,000 predicted stable ones, then about 800, 150 and 18 top candidates. PNNL synthesized one new solid electrolyte that uses approximately 70% less lithium and built a working battery from it; the end-to-end process took less than nine months.
Microsoft Discovery agents identify a non-PFAS coolant prototype in about 200 hoursMay 19, 2025. Microsoft said its researchers used the Microsoft Discovery agentic platform to find a novel non-PFAS immersion coolant prototype for data centres in about 200 hours, work it says usually takes months or years. The material was then synthesized in under four months and early property tests matched the AI predictions; it is still being analysed.
UChicago active-learning model finds four battery electrolytes from 58 data pointsSeptember 25, 2025. A University of Chicago team reported in Nature Communications an active-learning model that started from 58 data points and explored a virtual space of one million electrolytes. Over seven campaigns of about 10 electrolytes each, with real cells built and cycled, it identified four new electrolyte solvents that rival state-of-the-art electrolytes.
Discovery Learning predicts battery life from early tests2026 publication. Discovery Learning predicts battery cycle life from limited early experiments, including transfer from historical small-cell data toward large-format pouch cells.

Read the dated findings and sources

Status: Experiment-efficient battery prediction; validation-cost metric incomplete. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

2026 publication: Discovery Learning predicts battery cycle life from limited early experiments, including transfer from historical small-cell data toward large-format pouch cells. Discovery Learning battery lifetime prediction, Nature.

Interpretation and tracking value

Predicting lifetime faster is not the same as experimentally confirming a longer-lived battery. Total staff, compute and external-validation cost remain necessary.

Next comparable measurement

Count early tests and full-lifetime validation cycles, elapsed days and costs; evaluate prediction error on independently held-out chemistries and duty cycles. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Storage performance

35%estimated progress

What we track

Measured round-trip efficiency, useful cycle life, and lifetime cost per delivered kWh under the same duty cycle.

Why this estimate?

Adaptive cycling yields measured laboratory lifetime gains with feedback control. Commercial-scale efficiency, lifetime cost and autonomous development across duty cycles are not established.

Advancements

SLAC-Stanford: ML-guided fast first charge extends battery life about 50%August 29, 2024. A SLAC-Stanford Battery Center study in Joule found that giving lithium-ion cells their first (formation) charge at high current made test batteries last about 50% longer on average while cutting the first charge from about 10 hours to about 20 minutes. Machine learning identified temperature and charging current as the two factors that mattered most among dozens.
AI-managed cycling extends solid-state battery life 2.65×December 16, 2025. At 80% state of health, intelligent cycling averaged 1,915 cycles versus 723 for fixed cycling: approximately 2.65 times the cycle count.
Chalmers RL charging strategy extends simulated EV battery life by 22.9%12 May 2026. Chalmers University researchers reported a reinforcement-learning fast-charging controller that extended battery lifetime by 22.9% in equivalent full cycles to 80% state of health versus standard charging, with average charging time of 24.12 minutes against 24.15. The lifetime result comes from a high-fidelity simulation; the voltage-to-health mapping was validated on three-electrode cells and testing on physical batteries is the stated next step.

Read the dated findings and sources

Status: Measured laboratory cycle-life improvement. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

December 16, 2025: At 80% state of health, intelligent cycling averaged 1,915 cycles versus 723 for fixed cycling: approximately 2.65 times the cycle count. Real-time AI for solid-state lithium metal batteries.

Interpretation and tracking value

The proof of concept uses small cells and adaptive charging at high rates. It does not establish matched-duty-cycle grid-storage cost or round-trip efficiency at commercial scale.

Next comparable measurement

Replicate on commercial-format cells at matched delivered energy and duty cycle; measure efficiency, useful life and full lifetime dollars/kWh. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note