Skip to content

carryer123/whitlockite-ml

Repository files navigation

Ni-Whitlockite: Computational Screening (MD → Physics Score → XGBoost)

Method Model MD DFT Paper

The computational pipeline behind the PAiCER manuscript ranks nickel (Ni) as the top divalent substituent for metal-substituted whitlockite (a calcium-phosphate bone-graft biomaterial) — Rank #1 by this computational screen after the stated chemical filters. (The experimental confirmation of Ni's superior reinforcement is reported in the manuscript, not in this repository.) The screen runs molecular dynamics → a physics-based scoring function → an XGBoost extension over the full candidate library, and ranks Ni at #1 with a Physics Score of 282.19.

Scientific question. Which substituent M gives the best reinforcement profile in metal-substituted whitlockite? Hypothesis. A d8 transition metal (Ni) maximizes strength through crystal-field stabilization (CFSE) and a regular octahedral preference in a distorted lattice.

This repository contains the computational/ML pipeline, its input data, and reproduced result tables only. Experimental characterization (XRD / XPS / FTIR / nanoindentation) is reported in the manuscript and its Supporting Information.


Pipeline

Layer 1  Physics Score (21 metals, MD+theory) + XGBoost to 45 elems -> Table S1; Ni = 282.19 (Rank #1)
Layer 2  Weighted ranking of the 21 candidates (biocompatibility 20%) -> Table S2; Ni = 97.29 (Rank #1)
Verify   DFT (Ni Jahn-Teller-inactivity)                             -> CFSE / electronic origin

The screen is two-layer (PAiCER SI): Layer 1 is a broad physics-only Physics Score over a 5D Pareto front (Table S1, 0-300); Layer 2 is the biomaterial-facing weighted multi-objective ranking of the Pareto-viable candidates that adds LD50/IC50 biocompatibility as an explicit objective (Table S2, 0-100). Ni is Rank #1 in both.

2. Physics Score (21 metals)

src/physics_score_21metals.py computes the manuscript Physics Score for the 21 MD-screened metal-substituted whitlockites from MD- and theory-derived descriptors only. No experimental data is used.

The score is a weighted additive combination of:

group descriptors weight basis
dynamic stability MSD (50 ns, temporal stability, trend) physical, fixed
structural uniformity Lindemann CV, coordination-number CV physical, fixed
bond strength RDF peak height / width, time consistency physical, fixed
crystallinity calculated XRD peak count physical, fixed
electronic / geometric CFSE, ionic-radius match, Jahn-Teller penalty ligand-field theory

Running it reproduces Ni Physics Score = 282.1932 (Rank #1), matching the manuscript Table S1 (MD-data rows).

3. XGBoost extension to the candidate library

src/predict_candidates_xgboost.py trains an XGBoost regressor (Chen & Guestrin, 2016) on the 21 Stage-1 Physics Scores using theoretical atomic descriptors as features, then predicts the Physics Score for every candidate element. Ni remains Rank #1 (282.19); CFSE is the dominant feature. (The candidate library spans 62 elements across the periodic table; the manuscript Table S1 reports the 45 divalent-relevant candidates — Ni is Rank #1 in both.)

The authoritative reproduced ranking is outputs/physics_ranking_xgboost.csv (manuscript Table S1).

Note. The ranking uses XGBoost (Chen & Guestrin, 2016), the algorithm named in the manuscript. The Stage-1 Physics Scores for the 21 MD-screened metals — including Ni = 282.19 (Rank #1) — are model-independent and reproduce exactly.

4. Table S2 — weighted multi-objective ranking (Layer 2)

src/physics_score_S2_21candidates.py ranks the 21 Pareto-viable candidates by a weighted sum of five normalized objectives — the manuscript's biomaterial-facing final ranking (Fig. 1):

objective weight direction
Hardness 40 higher better (min-max)
CFSE 30 higher better
Biocompatibility (LD50/IC50) 20 lower toxicity better (inverted)
Lindemann CV 5 lower better (inverted)
dUtopia 5 pre-normalized Pareto-utopia distance

This is where biocompatibility enters as an explicit 20% objective, so physically-strong but unsafe candidates (e.g. Pb, Ra, Ba) fall in the ranking. The per-element objective values are the published PAiCER SI Table S2 values (data/objectives_21candidates.csv). Running the script reproduces Ni = 97.2881 (Rank #1) and matches all 21 published Weighted sums to within 1e-3 (expected/TableS2_21.csv). The manuscript synthesizes and characterizes the top-ranked viable set Ni, Co, Cu, Mg.

5. DFT verification

  • DFT: Quantum ESPRESSO (PBE+U) + LOBSTER address the electronic origin (CFSE / COHP) and confirm Ni's Jahn-Teller-inactive character (the proposed amorphization mechanism). A 42-atom primitive cell with a two-step protocol (nspin=1 relaxation → nspin=2 energy) handles Ni's near-degenerate high/low-spin states. See docs/DFT_convergence_criteria.md.

Experimental nanoindentation (which independently confirms the Ni system is by far the hardest) is reported in the manuscript; the ML ranking here does not take any measured hardness as input.


Reproduce

pip install -r requirements.txt
python src/physics_score_21metals.py          # Layer 1 -> outputs/physics_score_21metals.csv  (Ni = 282.1932)
python src/predict_candidates_xgboost.py       # Layer 1 -> outputs/physics_ranking_xgboost.csv (Table S1, Ni #1)
python src/physics_score_S2_21candidates.py    # Layer 2 -> outputs/physics_score_S2_21candidates.csv (Table S2, Ni = 97.2881)

expected/TableS1_45.csv and expected/TableS2_21.csv are the published SI answer keys used for verification.

Repository structure

src/        Layer-1 Physics Score + XGBoost extension; Layer-2 weighted ranking
data/       MD descriptors (21 metals), calculated XRD peaks, Table S2 objective values
expected/   published SI answer keys (Table S1 45-elem, Table S2 21-cand)
docs/       analysis & validation notes, DFT convergence, data dictionary
outputs/    reproduced result tables

Authors & Contributors

  • Jung Heon Lee@juhelee7 — supervision, corresponding author
  • Jina Bae@jinjin-del — experiments (synthesis & characterization)
  • Byoungsang Lee@carryer123 — MD / DFT / ML computation (this repo)

See CONTRIBUTORS.md.

Software

GROMACS (MD); Quantum ESPRESSO 7.3.1 + LOBSTER 5.1.0 (DFT); XGBoost, NumPy, pandas (ML).

Citation

Manuscript: PAiCER (Bae et al.), submitted to Advanced Materials. ML method: Chen, T. & Guestrin, C. XGBoost: A Scalable Tree Boosting System. KDD 2016.

About

XGBoost-reproducible ML pipeline for metal-substituted whitlockite biomaterials: in-house MD descriptors (MSD / Lindemann / PMF) -> 5D Pareto screening -> Ni optimum. CFSE-dominant hardness model (Advanced Materials submission).

Topics

Resources

License

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages