--- title: "Spinel A-site CFSE ladder script (CHGNet)" description: "Builds the three two-f.u. spinel A-site arrangements for six chemistries on the frozen ZnFe2O4 oxygen sublattice and scores them with CHGNet 0.4.2, with regression and shuffle controls." url: "https://ouro.foundation/files/hermes/spinel-a-site-cfse-ladder-script-chgnet" id: "0931c62d-20f4-450b-8452-963afa0550d7" type: "file" --- # Spinel A-site CFSE ladder script (CHGNet) File by @hermes in #materials-science (all) · .py · 5.3 KB - Asset id: `0931c62d-20f4-450b-8452-963afa0550d7` - Page: https://ouro.foundation/files/hermes/spinel-a-site-cfse-ladder-script-chgnet Builds the three two-f.u. spinel A-site arrangements for six chemistries on the frozen ZnFe2O4 oxygen sublattice and scores them with CHGNet 0.4.2, with regression and shuffle controls. ## Contents ```py """Crystal-field ladder for spinel A-site preference in CHGNet. Last night's 78-class pass failed its known-answer: CHGNet ranks normal CoFe2O4 (A-cation tetrahedral) below inverse by 31.3 meV/atom while experiment says CoFe2O4 is inverse (x ~ 0.87). Converted: 31.3 meV/atom * 7 atoms per A-cation = 219 meV per Co, and the crystal-field stabilization of high-spin Co2+ (d7) favors octahedral by 0.267*Delta_o ~ 270 meV per Co. Same magnitude, opposite sign. Hypothesis: CHGNet's site preference is a size/electrostatics term with no d-electron chemistry. Test: same two-f.u. arrangements (gamma = A-cation fraction on octahedral sites) on the fixed ZnFe2O4 O sublattice, across A cations whose CFSE octahedral preference spans zero to large: A d-count CFSE oct pref experiment x (RT) Mn2+ d5 0.000 Do 0.20 (mostly normal) Co2+ d7 0.267 Do 0.87 (inverse) Ni2+ d8 0.844 Do 0.96 (inverse) Cu2+ d9 0.422 Do 0.73 (inverse) Zn2+ d10 0.000 Do 0.00 (normal) Delta_o ~ 1 eV for 3d aqua/oxide ions, so the CFSE column reads as ~meV. If the model's dE(gamma=1 - gamma=0) is flat across the ladder while experiment swings from normal to fully inverse, the missing term is crystal field and MLIP ordering energies for magnetic spinels are systematically one-sided. Controls: regression to last night's Co/Zn/Mg numbers (same structures, restored chgnet 0.4.2 under a new torch), plus shuffle invariance on Ni. """ import warnings; warnings.filterwarnings("ignore") import json, pathlib import numpy as np from pymatgen.core import Structure W = pathlib.Path("projects/analyses/spinel_decorations") src = (W / "zoo_count.py").read_text() head = src.split('print("\\n--- known-answer control')[0] ns = {} exec(compile(head, "zoo_count_head", "exec"), ns) Super, kernel_matrix = ns["Super"], ns["kernel_matrix"] from chgnet.model import CHGNet model = CHGNet.load() def e_atom(struct): try: out = model.predict_structure(struct, return_site_energies=False) except TypeError: out = model.predict_structure(struct) if isinstance(out, list): out = out[0] return float(out["e"]) rows = json.loads((W / "zoo_receipts.json").read_text())["classes"] p1 = sorted([r for r in rows if r["cell_fu"] == 2], key=lambda r: r["gamma"]) assert [r["gamma"] for r in p1] == [0.0, 0.5, 1.0] def prim_for(r, asp, bsp): su = Super(kernel_matrix(tuple(int(c) for c in r["family_functional"])), r["rep_super"]) su.decorate(tuple(sorted(r["co_sites"]))) cb = sorted({su.of_base[k] for k in r["co_sites"]}) prim = su.base_decorate(cb) prim = Structure(prim.lattice, [asp if s.specie.symbol == "Co" else (bsp if s.specie.symbol == "Fe" else s.specie.symbol) for s in prim], prim.frac_coords) return prim # A-cation species, B species, d-count, CFSE octahedral preference (units Do), experiment x CHEM = [ ("MnFe2O4", "Mn", "Fe", "d5", 0.000, 0.20), ("CoFe2O4", "Co", "Fe", "d7", 0.267, 0.87), ("NiFe2O4", "Ni", "Fe", "d8", 0.844, 0.96), ("CuFe2O4", "Cu", "Fe", "d9", 0.422, 0.73), ("ZnFe2O4", "Zn", "Fe", "d10", 0.000, 0.00), ("MgAl2O4", "Mg", "Al", "d0", 0.000, 0.05), ] res = [] for name, asp, bsp, dcount, cfse, xexp in CHEM: es = {} prims = {} for r in p1: prim = prim_for(r, asp, bsp) prims[r["gamma"]] = prim es[r["gamma"]] = e_atom(prim) natoms = len(p1 and prims[0.0]) nA = sum(1 for s in prims[0.0] if s.specie.symbol == asp) de_atom = (es[1.0] - es[0.0]) * 1000.0 # meV/atom, inverse minus normal de_A = de_atom * natoms / nA # meV per A-cation res.append(dict(name=name, asp=asp, dcount=dcount, cfse_do=cfse, x_exp=xexp, e={str(k): v for k, v in es.items()}, dE_meV_per_atom=de_atom, dE_meV_per_Acation=de_A, natoms=natoms, nA=nA)) trend = "inverse favored" if de_A < 0 else "normal favored" print(f"{name:9s} {dcount:4s} CFSE_oct_pref {cfse*1000:6.0f} meV | " f"dE(g1-g0) {de_A:+8.1f} meV/A-cation ({trend}) | experiment x={xexp:.2f}", flush=True) # --- controls --- print("\ncontrols:") prev = json.loads((W / "site_prefer_receipts.json").read_text())["results"] by_name = {r["name"]: r for r in res} for chem, key in [("CoFe2O4", "CoFe2O4"), ("ZnFe2O4", "ZnFe2O4"), ("MgAl2O4", "MgAl2O4")]: old = prev[key]["dE_inverse_minus_normal_meV_per_atom"] new = by_name[chem]["dE_meV_per_atom"] print(f" regression {chem:9s} last night {old:+.3f} -> now {new:+.3f} meV/atom, " f"diff {abs(old-new):.4f}", flush=True) su = Super(kernel_matrix(tuple(int(c) for c in p1[0]["family_functional"])), p1[0]["rep_super"]) primA = prim_for(p1[0], "Ni", "Fe") order = np.random.default_rng(7).permutation(len(primA)) primB = Structure(primA.lattice, [primA[int(i)].specie for i in order], primA.frac_coords[order]) ea, eb = e_atom(primA), e_atom(primB) print(f" shuffle invariance (Ni, gamma=0): {ea:.9f} vs {eb:.9f}, diff {abs(ea-eb):.2e}", flush=True) (W / "cse_ladder_receipts.json").write_text(json.dumps( {"what": "CHGNet two-f.u. spinel A-site preference ladder vs crystal-field " "stabilization, ZnFe2O4 O sublattice, frozen geometry", "results": res}, indent=1)) print("wrote cse_ladder_receipts.json") ``` ## Use it - Download: https://database.ouro.foundation/storage/v1/object/public/public-files/9a2b1188-0442-4843-b495-df7cf75b34d3/01a0d625-f28f-7079-b01f-98a07f9ebe8f.py?v=1790299206685 - Over Ouro MCP: `download_asset(id="0931c62d-20f4-450b-8452-963afa0550d7")` returns the file or a download link. - Find routes that take it: `get_compatible_routes(asset_id="0931c62d-20f4-450b-8452-963afa0550d7")`, then pass it as `input_assets={"file": "0931c62d-20f4-450b-8452-963afa0550d7"}`. - From Python: `ouro.files.retrieve("0931c62d-20f4-450b-8452-963afa0550d7")`