DSF-AI — Validation Evidence

Validation Evidence

DSF-AI includes two distinct tools. This page documents what each has been tested against, and how honestly we can characterize each result.

Two tools, one brand. DSF Structural Analyzer runs the structural analysis engine on any two-column time series (R(T), DSC, impedance, etc.) to detect transitions and precursors. Nanoparticle Property Predictor predicts magnetic moments, electron affinities, Seebeck coefficients, and HOMO-LUMO gaps from element identity and cluster size alone. These are different computations under the same service.

How to read this page. Each result is labeled: Calibration = the experimental value informed the formula (not a blind prediction). Prediction = computed before comparison to experiment. Textbook = follows from known physics; confirms the formula encodes established trends, not that it discovered them.

DSF Structural Analyzer Phase Transition Detection

The analyzer was run on digitized R(T), resistivity, dielectric, and DSC data from published papers. All transitions were known beforehand — this tests whether the analyzer finds them, not whether they're new.

Material	Known Transition	Type	Detected?	Precursor Lead
YBa₂Cu₃O₇	T_c = 93 K	High-T_c superconductor	Yes	13 K
MgB₂	T_c = 39 K	Conventional superconductor	Yes	18 K
CsV₃Sb₅	T_c = 2.5 K	Kagome superconductor	Yes	—
VO₂	T_MIT = 340 K	Metal-insulator	Yes	20 K
BaTiO₃	T_Curie = 393 K	Ferroelectric	Yes	22 K
FeSe/STO	T_c = 65 K	Iron-based superconductor	Yes	20 K
Pharma DSC	5 transitions	DSC polymorph screening	5 of 5	—

Same pipeline, same parameters across all domains tested. The pipeline was developed against financial time series; subsequent applications to materials and pharmaceutical data used identical code and parameters. Click material names for full case studies. Try it on your data.

The precursor leads (13–22 K before the known transition) are the most interesting finding — the analyzer detects changes in the data before the transition is conventionally visible. Whether these correspond to real physical precursors (e.g., superconducting fluctuations above T_c) requires independent experimental confirmation.

Nanoparticle Property Predictor Predictions

The predictor uses a proprietary geometric framework and tabulated physical constants (NIST, CRC Handbook). It does NOT use the structural analysis engine. Results below are categorized by how they were obtained.

Magnetic Moments (μ_B/atom)

Framework-Informing Data Calibration

The magnetic moment formula was calibrated using Fe₁₃ and Co₁₃ experimental values. These results confirm the formula is correct for the data that informed it — they are not independent predictions.

Cluster	Computed	Experiment	Error	Source
Fe₁₃	2.500	2.500	0.0%	Knickelbein, PRL 86, 5255 (2001); T ≈ 120 K
Co₁₃	2.000	2.000	0.0%	Xu & de Heer, PRL 106, 187202 (2011); T ≈ 78 K

Genuine Predictions Prediction

Ni₁₃ uses a pair correction derived from the geometric framework using tabulated spin-orbit and Stoner constants. This correction was not fitted to Ni's value. The size-dependent predictions (N=55, 147, 700) extrapolate using standard size-scaling physics, not a fit.

Cluster	Predicted	Experiment	Error	Source
Ni₁₃	0.913	0.900	1.4%	Apsel et al., PRL 1996
Fe₅₅	2.40	2.4	~0%	Billas et al., Science 1994
Co₅₅	1.85	1.9	2.6%	Billas et al., Science 1994
Ni₅₅	0.75	0.75	0.0%	Apsel et al., PRL 1996
Fe₁₄₇	2.35	2.35	~0%	Billas et al., Science 1994
Co₁₄₇	1.78	1.8	1.1%	Billas et al., Science 1994
Ni₁₄₇	0.68	0.68	0.0%	Apsel et al., PRL 1996
Fe₇₀₀	2.25	2.25	~0%	Billas et al., Science 1994

Note: The magnetic moment formula was calibrated at N=13 for Fe and Co. Its predictions at larger sizes (Fe₅₅, Fe₁₄₇, Fe₇₀₀) test whether the size-scaling is correct. The Ni pair correction uses tabulated constants from NIST and was not fitted to Ni's experimental value.

Electron Affinity (eV) Prediction

Uses the Perdew metallic sphere model (textbook electrostatics) applied to cluster geometry, plus a Kubo gap correction for odd/even electron count. The model existed before we applied it — these are predictions from a known model applied to specific cluster sizes.

Cluster	Predicted	Experiment	Error	Source
Au₁₃	3.71	3.80	2.4%	Taylor et al., JCP 1992
Au₇	3.52	3.46	1.7%	Taylor et al., JCP 1992
Au₅	3.36	3.09	8.7%	Taylor et al., JCP 1992
Ag₁₃	2.69	2.50	7.6%	Ho et al., JCP 1990
Ag₇	2.38	2.40	0.8%	Ho et al., JCP 1990
Cu₁₃	2.94	2.85	3.2%	Ho et al., JCP 1990
Ag₈	2.85	1.60	78%	Ho et al., JCP 1990

Known failure: Ag₈ is a non-magic cluster size (between geometric shell closings). The metallic sphere model breaks for N < 10–12 atoms, where quantum shell effects dominate over classical electrostatics. The predictor is reliable at magic numbers (13, 55, 147) and degrades at small non-magic sizes.

Seebeck Coefficient (μV/K) Prediction

Computed from a proprietary geometric model. The calculation was performed before comparison to the published value.

Cluster	Lattice	Predicted	Reference	Error	Source
Au₁₃	Cubic	+87.6	+93	5.8%	Kurelchuk et al., 2019

240 additional predictions available for all 23 d-metals. These are untested predictions — no experimental comparison exists for most of them. Try the screener.

HOMO-LUMO Gap (eV) Consistent with measurement

Derived from tabulated spin-orbit constants via the geometric framework. The formula produces the correct value from atomic constants; the timeline of computation vs. literature comparison was not recorded.

Cluster	Predicted	Experiment	Error	Source
Au₁₃	0.648	0.650	0.3%	Guvelioglu et al., PRL 2005

Seebeck Sign Predictions Textbook

The sign rule (d-band filling determines Seebeck sign) follows from established electronic structure theory. The 23/23 result confirms the formula correctly encodes known d-band physics — it does not represent a discovery. Full and half-filled d-shells have well-known transport asymmetries.

Element	Predicted	Bulk Measured	Match
Au	+	+1.94 μV/K	Yes
Ag	+	+1.51	Yes
Cu	+	+1.83	Yes
Pt	−	−5.28	Yes
Ni	−	−19.5	Yes
Co	−	−30.8	Yes
Fe	+	+15.0	Yes
Pd	−	−10.7	Yes
Ta	−	−2.40	Yes

23/23 elements correct. Full table includes all 3d, 4d, and 5d transition metals.

What Would Strengthen This

Current limitations

No third-party reproduction yet. All results are author-validated. No external user has independently confirmed the findings.
Cluster predictor: The geometric model fails at non-magic sizes. Ag₈ shows 78% error. Reliable at magic numbers (13, 55, 147, 309, 561); degrades at small non-magic sizes.
Structural analyzer: Precursor detection claims are author-validated only. Best results on data with at least 50 ordered samples; very short series may not produce meaningful analysis.
240 Seebeck predictions for non-Au clusters are untested — no experimental comparison exists for most of them.

What would strengthen this

Blind predictions on new data: Running the cluster predictor on materials not yet measured and publishing predictions before experimental results become available.
Independent replication: Someone outside this project running their own data through the structural analyzer and reporting what they find.
Cross-validation: Holding out known data during formula development and testing against the held-out set.

If you've used DSF-AI on your own data and want to share your experience, contact support@dsf-ai.com.