chrisroadmap
commented
1 year ago After a little more digging, this is only triggered with certain combinations of stochastic parameters in the energy balance model. Prior to scipy 1.10, it was only a warning, and now it is an error. It seems to be tricky with non-annual time steps and stochastic, this should be treated with care.
Another point is that the calibration (up until now) is done in R, and the model run in Python. It seems like R's linear algebra routines are a little closer to Python's (specifically, scipy's) sparse linear algebra rather than the full linear algebra package. The CMIP6 calibration done in R will sometimes throw these not positive semi-definite errors with the full linear algebra, but the calibrations seem to work OK with the sparse algebra. So for now I'm retaining the sparse algebra in the energy balance model, but at some point it should be replaced. This will require a recalibration and will give different results, so it is a target for a minor version release.
In the calibration: need a pre-filter to eliminate combinations of non-positive semidefinite. Could be a lot of combinations filtered out.
ValueError: The input matrix must be symmetric positive semidefiniteis raised with scipy 1.10.0, and did not occur with scipy 1.9.3.See https://scipy.github.io/devdocs/release.1.10.0.html under "Other changes".