[MarcBerliner]

Description

Adds the IDAKLUSolver option upsampling_factor. When it is above 1.0, the solver saves extra points inside each accepted step until linear interpolation of one scalar output matches the integrator's own interpolant to tol / (factor - 1), where tol = atol + rtol·|v|. So 1 is off, 2 is about the solver tolerance, and larger values add more points. Points are only ever added.

How it works (inline, in the C++ step loop)

• The scalar variable is generated with the existing process() codegen, the same path as output_variables, and passed to the solver group as upsampling_fcn(t, y, p).
• After each accepted step, the value at t_val is evaluated once from the state already loaded and reused as the next step's start. The value at the step midpoint comes from IDAGetDky, written into y_cache. If |(v_a + v_b)/2 − v_m| misses the budget, the step is bisected and each midpoint is saved through the normal SavePoint.
• No derivatives are used. It works with or without stored Hermite data, with output_variables, and with sensitivities. Sensitivities at added points are IDA's own interpolated values.
• Experiment steps choose the variable through upsampling_variable: "Voltage [V]" by default, "Current [A]" for voltage steps. It is attached to the built model at build time. The unified model keeps the voltage default.
• Algebraic states are skipped, because IDA's error test already covers them. In a voltage hold the current is the algebraic state, so upsampling does nothing in that step.
• It is mutually exclusive with hermite_reduction_factor > 1, and inactive when t_interp is given.

SPMe (Chen2020) benchmarks, fastest of 15 warm solves:

config	1C discharge	5-step experiment	+ 2 sensitivities
off	2.2 ms, 99 pts	17.5 ms, 407 pts	33.5 ms
factor 1.1	+0.9 ms, 110 pts	+2.6 ms, 420 pts	+2.7 ms
factor 2	+1.2 ms, 148 pts	+4.3 ms, 505 pts	+4.5 ms
factor 21	+4.0 ms, 443 pts	+11.8 ms, 1383 pts	+14.6 ms

Max linear-interpolation error of V between saved points, relative to the solver tolerance, on the 1C discharge: 350 (off), 13 (1.1), 2.5 (2), 0.42 (5), 0.08 (21).

A cubic Hermite version that also used the analytic dV/dt was explored first (5e86f72e4) and replaced. It needed fewer points for the same error, but cost about twice as much per solve at loose budgets.

create_casadi_solver_group gains a required upsampling_fcn argument, so this needs the matching pybammsolvers build.

Type of change

• New feature: CHANGELOG entry under ## Features.

Important checks:

• No style issues: pre-commit run on changed files
• Tests pass locally: packages/pybamm/tests/unit/test_solvers, test_experiments and test_simulation.py (932 passed); the pybammsolvers tests (70 passed)
• The documentation builds: nox -s doctests (not run)
• Tests added: TestUpsampling in test_idaklu_solver.py, and test_upsampling_variable in test_base_step.py

[github-actions]

Bencher Report

Project	PyBaMM
Branch	claude/linear-upsampling-experiment
Testbed	bare-metal

⚠️ WARNING: Truncated view!

The full continuous benchmarking report exceeds the maximum length allowed on this platform.

🐰 View full continuous benchmarking report in Bencher

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.23%. Comparing base (5b81c32) to head (2ef8b64).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #5998   +/-   ##
=======================================
  Coverage   98.23%   98.23%           
=======================================
  Files         350      350           
  Lines       36104    36127   +23     
=======================================
+ Hits        35468    35491   +23     
  Misses        636      636           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.