Research

My research asks how geochemical, geological, spatial, and remote-sensing evidence can be converted into robust mineral-exploration inference without losing geological meaning.

The objective is not simply to maximize predictive performance. I am interested in what information the data genuinely contain, how that information transfers across deposits and regions, how uncertainty should be represented, and when a model result is geologically defensible enough to support a decision.

Core research questions

Mineral chemistry as geological information — Which trace-element and isotopic signatures in pyrite, galena, sphalerite, and related minerals are reproducibly associated with ore system, mineralization style, tectonic environment, or critical-metal enrichment?
GeoAI under geological dependence — How should machine-learning models be validated when samples are spatially clustered, deposit-dependent, imbalanced, or generated by heterogeneous analytical programs?
Anomaly to target — How can multivariate geochemical anomalies be distinguished from statistical outliers and translated into spatially coherent exploration evidence?
Multimodal prospectivity — How should geology, geochemistry, geophysics, structures, terrain, remote sensing, field observations, drilling, and uncertainty be integrated without hiding provenance or geological contradictions?
Uncertainty-aware exploration decisions — How can geostatistics, sensitivity analysis, prospective testing, and model auditability improve the reliability of target ranking and exploration recommendations?

Selected research evidence

5,200+ analysesGlobal pyrite dataset spanning 138 deposits and multiple ore-system classes.

AUC > 0.99Reported in selected pyrite-discrimination workflows under study-specific validation.

37 Pb-Zn depositsGalena geochemistry evaluated with RF, GB, MLP, and SVM workflows.

43 global locationsSedimentary-pyrite tectonic-setting research using CNN/XGBoost and interpretable analysis.

Performance values are study-specific validation results, not universal model guarantees. Interpretation depends on cohort definition, validation design, label quality, feature availability, analytical comparability, and transfer domain.

Research programs

Program 1 — Mineral geochemistry & metallogenic discrimination

Scientific objective: test whether mineral-scale geochemistry retains reproducible information about ore-forming system and geological setting.

Research uses global and deposit-scale pyrite, galena, and sphalerite data, including LA-ICP-MS trace elements, S-Pb isotopes, geochemical ratios, geological labels, imbalance treatment, and deposit-aware evaluation.

Representative outputs include:

global pyrite metallogenic typing across multiple ore-system classes;
galena-based Pb-Zn metallogenic discrimination across 37 deposits;
pyrite-based Pb-Zn discrimination and geological-class testing at Gunga;
sphalerite research on mineralization zones, ore genesis, and critical-metal enrichment.

Selected publications · Public research repositories

Program 2 — GeoAI & interpretable scientific machine learning

Scientific objective: determine when predictive structure is robust enough to support geological inference or exploration decisions.

Methods include Random Forest, XGBoost, SVM, gradient boosting, MLP/CNN workflows, clustering, PCA, t-SNE/UMAP, SHAP, LIME, imbalance treatment, grouped validation, and related methods where scientifically justified.

The emphasis is on validation architecture, data leakage control, transferability, uncertainty, and geological interpretability, not algorithm novelty by itself.

Program 3 — Geochemical anomaly intelligence

Scientific objective: distinguish multivariate exploration signal from background variability, analytical artifacts, and isolated statistical extremes.

Research and professional R&D include correlation structure, robust outlier diagnostics, clustering, manifold learning, Isolation Forest-style scoring, threshold sensitivity, spatial confidence, sample-type effects, and target-priority ranking.

This program connects directly to my professional work on GeoAnomalyAI and related exploration-geochemistry workflows, while proprietary implementation details remain non-public.

Program 4 — Mineral prospectivity & multimodal evidence integration

Scientific objective: combine heterogeneous exploration evidence while preserving provenance, geological meaning, and contradictions between datasets.

Inputs can include geology, geochemistry, geophysics, structures, terrain, remote sensing, alteration evidence, field samples, trenches, drilling, and uncertainty. The goal is not a single opaque score, but a defensible chain from source evidence to prospect screening, confidence assessment, and target prioritization.

Program 5 — Geostatistics & spatial uncertainty

Scientific objective: represent spatial continuity and uncertainty explicitly rather than treating interpolated values or model scores as deterministic truth.

Work includes variography, kriging, sequential Gaussian simulation, uncertainty products, spatial validation, target sensitivity, and GPU/CUDA-oriented experimentation for scalable simulation workflows.

Validation and falsification principles

For geological and spatial ML, random row-wise train/test splitting can materially overstate generalization when samples share deposit, campaign, spatial, or analytical context. My preferred controls therefore include:

group-aware or deposit-aware validation when observations are clustered by site or deposit;
spatial or geological holdouts where random splitting would leak context;
strict separation of training, feature selection, threshold selection, calibration, and holdout evaluation;
sensitivity analysis for thresholds, cohorts, missing data, outliers, class balance, and feature choices;
explicit checks for label leakage, duplicate analyses, site-level dependence, and circular geological definitions;
external or prospective validation where feasible;
explicit separation of prediction, association, interpretation, and causal claims;
reporting of failure, instability, non-transferability, and negative results.

From prediction to geological interpretation

I do not treat feature importance, SHAP values, embeddings, or high classification accuracy as direct evidence of geological mechanism. These outputs can identify patterns worth investigating, but mechanistic interpretation requires independent geological, mineralogical, geochemical, spatial, and process-based evidence.

This distinction is central to my work: prediction is a computational result; geological interpretation is a scientific argument.

Computational and geoscience methods

Research outputs and current work

Publications — selected peer-reviewed, online-first, and current manuscript outputs.
Projects — public research repositories plus capability-level descriptions of professional R&D.
GitHub — intentionally released code and research assets.
AIMEX Lab — independent GeoAI research and benchmarking initiative.