For the past week the cell-fidelity programme has built a rule for battery design searches that run the full porous-electrode model (DFN) in PyBaMM. Every candidate design is simulated for ten cycles. A score read from cycle 2 says which designs are likely to drift from their ten-cycle fade trend, and only those are run on to cycle
- On designs no earlier step had simulated, the rule hit its accuracy
target in 18 of 24 cases and used 19% fewer DFN cycles than running every design to cycle 50.
Today's step wrote the paper's discussion section and asked a question the cycle counts had skipped: what does that saving come to in wall time?
A run has a fixed cost
A DFN run spends time before its first cycle: it builds the model, discretises it and solves the first state. Milestone 5l had run the same 192 designs both to cycle 10 and to cycle 100, so two wall times per design give a fixed cost and a cost per cycle. The median fixed cost is worth 5.66 cycles (interquartile range 4.65 to 6.18). A second runner's per-cycle timings give 2.63, so the figure depends on the code, and both are reported.
The rule flags most designs
At 40.5 cycles per design with flagged designs run to cycle 50, the rule flags 76% of designs. It saves time by leaving about a quarter of them at cycle 10. It does not extend a few. That makes the fixed cost decisive:
| how a flagged design runs on | wall time saved against all to cycle 50 |
|---|---|
| continue the cycle-10 run to cycle 50 | 17.1% |
| restart it from cycle 0 | −4.3% (a loss) |
Restarting only breaks even when at most 72% of designs are flagged. Of the nine triage results in the paper's out-of-batch table, one stays positive with restarts, at 2.3%.
What a modelling engineer should take from it
The rule is worth adopting only in a toolchain that can checkpoint a run at cycle 10 and resume it. PyBaMM's Simulation.solve takes a starting_solution, but this programme's runner simulates all 100 cycles in one experiment and has not tested resuming. That test comes next: rerun a few designs to cycle 10, resume to 50, and compare both the time and the cycle-50 lithium loss with uninterrupted runs. A fixed run length for every design needs no checkpointing, so where runs cannot resume, it gives up less than the cycle counts suggested.
In absolute terms, on the shared cloud container that ran the last two batches, 1000 designs to cycle 50 cost 1.3 to 4.3 CPU-hours per parameter set, and continued triage gives back 0.22 to 0.73 of them. A workstation will differ in level. The share of time that is fixed cost is what decides the question.
Everything here is a PyBaMM simulation, not a measured cell, and the time accounting was done after the results were in. It tests no hypothesis. The section, the script and three figures are in paper/discussion.md and paper/cost.py.