Among the stable candidates, which balance oxidation, weight and brittleness best?

EXPERIMENTS.md · lines 6991–7027

E118 — Multi-objective selection over the feasible set

Oxidation, liquid-oxygen compatibility and cold brittleness are selection criteria, not feasibility ones: they are applied to whatever survived the ladder and they trade against one another. forager/selection.py implements them as a separate layer with the constants and encodings sourced in its docstring, and with the brittleness figure deliberately ordinal because the ORNL source states the transition "is not a fixed property of the material" and moves by 204 C on fabrication variables alone.

Of the 55 feasible compositions, 29 contain no LOX-prohibited element and 14 are on the Pareto front over (oxidation, density, cold brittleness, driving force):

 oxid  g/cm3  brittle    drive   composition
 0.06   7.62     1.50    -30.2   Cr0.75 Ni0.25
 0.18   8.33     0.67    +48.3   Co0.33 Cr0.33 Ni0.33
 0.26   8.49     0.50    +72.8   Co0.25 Cr0.25 Cu0.25 Ni0.25
 1.70  17.30     1.25    -32.8   Ta0.75 W0.25
 2.05  13.46     1.50    -86.6   Mo0.50 Ta0.50   <- the driving-force optimum
 2.88   7.15     1.25    -11.0   V0.75 Mo0.25

The composition that wins the ladder is near the worst thing on the front for the application. Mo0.50 Ta0.50 carries an oxidation penalty of 2.05 against Cr0.75 Ni0.25's 0.06, and is 5.8 g/cm^3 heavier. It is the right answer to "what holds a single phase" and a poor answer to "what survives oxygen at 1000 K".

This is the concrete payoff of the twelve-element widening (E114): every composition in the top three of this front is one the eight-element space could not represent. Fable predicted this before the run - "expect to rediscover the Monel/Inconel neighbourhood, which is a validation, not a failure" - and Cr-Ni and Co-Cr-Ni are that neighbourhood.

One caveat that may kill the leader, and our ladder cannot see it. Chromium has an antiferromagnetic Neel transition at 311 K, which is inside the 90-1000 K window. The expansion and MACE are both spin-blind, so a magnetic transition is invisible to every rung. Cr0.75 Ni0.25 is therefore reported as feasible by a model that cannot evaluate the most likely way it fails. Nothing Cr-rich should be promoted up the ladder until a magnetic check exists. The same applies to Ni (Curie 627 K) and Co (hcp-fcc at 695 K), both also inside the window.