[GSA] SCT + our screening approach

[aleksandra-kim]

• add std to mean to the heuristic function
• add var to mean to the heuristic function
• test both on average hh consumption
• refactor the code to be easily applicable for archetypes

[aleksandra-kim]

what seems to work for quick GSA:

1. bio and cf

• • local sa - 10 min max

2. tech

• • sct with very low cutoff 1-16 for tech - 30 min
• • out of these, select 2k with highest contributions
• • for these 2k run local sa - 30 min, no dask

3. select 4-5k tech, bio, cf inputs based on local SA

4. run validation, correlation should be very high, >0.999

5. compute spearman indices for 5k inputs - 2-3h with 20 dask workers, (can be much faster with more workers)

• • in our case spearman is NOT good enough for ranking

6. run validation for first 200 inputs

This has been tested on average CH HH, and ecoinvent 3.7.1

[aleksandra-kim]

Regarding heuristic function in the supply chain traversal (SCT) algorithm:

1. I tried various measures of uncertainties: std2mean, var2mean, 95% confidence range to mean for all exchanges from SCT with very low threshold (eg 1e-16)

• • normalized all of the them by respective max values
• • most appropriate seems to be std2mean or range2mean (very high correlation between these two, but std2mean is less effort)
• • var2mean was not good, because it's not unitless, so given that amounts of some exchanges are in the order of 1e-5, and the others - 1e+5, this becomes a problem
• • haven't tried using pedigree matrix values, which are specified for all uncertainty types. this could be good, because it allows comparison between different distributions, eg we can compute lognormal std for all of them and it's unitless. Big but: general inputs (parameters) would not have pedigree values

2. In the end exchanges that showed up as most uncertain were NOT among 200 identified with localSA, they were mostly landfill and waste related, not sure why.

• • if a heuristic function is created as weighted average of uncertainty measures (eg std2mean) and contribution scores, it would first favor those landfill exchanges that are not among influential ones in GSA.
• • unless other uncertainty measure is used, it doesn't make sense to use heuristic function
• • if heuristic function is used, contributions obtained with SCT should also be normalized by max value

3. TODO later: might be worth trying to sample during SCT