How much better can a lithium-ion cell get by redesigning its electrodes, and does a smarter search help?
research/energy-cell
How much better can a Li-ion cell get by redesigning its electrodes, and does a smarter search help? In a PyBaMM model, thicker, more porous electrodes with more negative capacity lift energy after 100 cycles from 109 to ~143 Wh/kg (+30%); random, Bayesian and session-agent proposers end within 3% of each other. One aggressively ageing parameter set. Stylised.
What it found
A PyBaMM single-particle model with electrolyte of an NMC811 / graphite-SiOx cell, scored on discharge energy per kilogram of electrode stack after 100 cycles at a fixed current. Four design numbers (positive thickness, two porosities, negative-to-positive capacity ratio) are searched with 30 evaluations by random search, Gaussian-process Bayesian optimisation (5 seeds each) and the session agent proposing by hand (one run). The reference cell scores 109 Wh/kg; all three reach 141–143 (+30%) with a thicker, more porous positive electrode and about 30% more negative capacity, and their seed ranges overlap, so no proposer is shown to be better. Limits: one aggressively ageing parameter set, constant-current charge only, stack not finished cell, optimum at the search bounds. Stylised and simulated, not a tested cell.
A figure from its own output
Paper, code and results
More in Energy supply and dense energy storage
- How robust is the cheapest energy-storage technology at each duration when costs are uncertain?How robust is the cheapest storage technology by duration to uncertain costs? Monte Carlo over
energy-lcos: Li-ion wins 96% of scenarios at 2 h but 18% at 8 h; no technology wins over 48% between 6 and 12 h; caverns win most from 24 h. Ranges assumed. Stylised. - Do the lithium-ion cell designs found with a fast model still win under the full porous-electrode model?Re-scoring 49 designs from energy-cell under PyBaMM's full DFN model instead of SPMe: rank correlation 0.86 at cycle 100, the SPMe winner is 3rd of 45 and still +38% over the reference cell (+31% under SPMe), DFN reports about 19 Wh/kg more on average, and SPMe fails to cycle 3 designs DFN handles (modelled, stack level)
- How hard does a least-cost zero-emission German grid cycle its storage, and does the levelised-cost map survive that use?How hard does a least-cost zero-emission German grid cycle its storage? Batteries about 1,700–1,900 h a year (near the assumed 2,000), cavern hydrogen only 719 h, which more than doubles its LCOS to $864/MWh.
- Which storage technology is cheapest at each duration once costs come from NREL ATB and PyPSA technology-data, not from round numbers?Which storage technology is cheapest at each duration once costs are sourced (NREL ATB 2024, PyPSA technology-data)? Li-ion to 4 h, cavern compressed air from 6 h; at 8 h pumped hydro 52%, air 27%, Li-ion 22% of scenarios; hydrogen's assumed 500 h win does not survive.
- How much of the cost of a zero-emission German power system is the storage assumption, once costs are sourced?What does a zero-emission German power system cost once
energy-mix's assumed costs are replaced by sourced ones? PyPSA on 2019 hourly demand with PyPSA technology-data costs (Danish Energy Agency et al.): EUR 222/MWh with a steel hydrogen tank, EUR 144 with caverns, EUR 75 unconstrained; the storage type moves the answer more than the cost year. - Which weather year sets the hydrogen store of a zero-emission German power system, and how much does the cost depend on it?Which weather year sets the hydrogen store of a zero-emission German power system? The same sourced-cost PyPSA model on 2015 to 2019: EUR 203 to 271/MWh with a steel tank, EUR 141 to 155 with caverns, and a cavern store of 24 to 36 TWh depending on the year.
All 39 Energy supply and dense energy storage projects →
Other research topics
Proper scoring & elicitation28Effort & peer prediction9Verification games & slashing19Detection & tolerances14Markets as routers13Prediction-market design9Paying contributors8Decentralised training systems23Digital twins & sim-to-real164LLM chip design loops40Robot manufacturing loops53Ambient AI assistant46Self-replicating manufacturing34Simulation-first design64Engineered clothing45Task-specific AI accelerators40Algorithmic chip-design loops42Open-source purpose-built robots44The software-engineering loop43Adversarial human imitation41Open-source residential construction41Autonomous multi-angle capture34Sensing, imaging and 3D31Energy supply and dense energy storage39Robot societies among people31Agentic co-design of low-cost robots27Laundry-hanging robots25Machine learning on qualitative data23Robots on the building site27Civilizations & group psychology1History & human sciences1
Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.