Banca de QUALIFICAÇÃO: ANDRÉ FELIPE MATIAS COSTA ALDECI

Uma banca de QUALIFICAÇÃO de MESTRADO foi cadastrada pelo programa.
STUDENT : ANDRÉ FELIPE MATIAS COSTA ALDECI
DATE: 30/09/2026
TIME: 14:30
LOCAL: Auditorio 2 do DIMAp
TITLE:

Dual-Driven DDPM-TGCNN Architecture:
Synthetic Generation and Prediction of Light Rare
Earth Occurrences

KEY WORDS:

Rare Earth elements, Deep Learning, mineral prospection mapping,
Explainable Artificial Intelligence, multiobective optimization.

PAGES: 119
BIG AREA: Ciências Exatas e da Terra
AREA: Ciência da Computação
SUBÁREA: Metodologia e Técnicas da Computação
SUMMARY:

The growing global demand for Rare Earth Elements (REEs) and strategic
metals drives the search for more efficient mineral exploration methods, given
the distributive complexity and the critical importance of these resources for
the energy transition. Although Mineral Prospectivity Mapping (MPM) is
a consolidated technique, the approach faces severe challenges, such as the
multimodal nature of geoscientific data, the scarcity of known mineralized tar-
gets, and the difficulty in modeling spatial uncertainty in undersampled areas.
To mitigate these limitations, this paper proposes a hybrid and dual-driven
predictive methodology, which integrates Artificial Intelligence with a priori
metallogenic knowledge. Applied to the Barreiro Alkaline-Carbonatite Complex
(Araxá/MG, Brazil), the developed model (DDPM GraphTransformer) fuses
airborne geophysical, lithological, and geochemical data through Graph Neural
Networks (GNNs) and Transformer architectures. The core methodological
innovation consists of using a Hybrid Topology (Neural Kriging), which recon-
ciles local spatial connections with long-range anchors structured on boreholes.
This architecture is coupled with a Denoising Diffusion Probabilistic Model
(DDPM) to perform geochemical inpainting, generating an ultra-high-resolution
grid with 15,069 predictive points and quantifying the exploratory risk. The
architecture calibration was conducted via Multiobjective Hyperparameter
Optimization (MOHPO) using the MOTPE (Multiobjective Tree-structured
Parzen Estimator) algorithm. The model found the utopian point between geo-
logical realism (measured by the Wasserstein Distance) and the minimization of
mathematical error, achieving significant convergence with the Cost-Sensitive
Focal Loss metric stabilized at [insert final loss value]. The learning process was
strongly directed towards the detection of Light Rare Earth Elements (LREEs)
and Lithium. The pipeline also integrates eXplainable AI (XAI) techniques
via DeepSHAP to rank the weight of geophysical signatures in the predictions.
The results attest to the exceptional accuracy of the stochastic ensemble. The
network accurately delineated severe anomalous halos, isolating prospective
targets with maximum grades of up to 469.60 ppm for Cerium and 274.80 ppm
for Lanthanum, in addition to recording a very low exploratory risk, with a
median standard deviation of only 0.10 ppm across the grid. The locational
accuracy of the model, rigorously validated by Prediction-Area (P-A) plots,
demonstrated exceptionally high predictive efficiency, achieving hit rates of
95% for Cerium (Ce) and Lanthanum (La), and 81% for Lithium (Li). This
algorithmic precision, combined with the application of geostatistical interpo-
lators (Ordinary Kriging and IDW) on the network’s output, proved highly
effective in delineating continuous metallogenic corridors. As future steps, the
methodology will be benchmarked against state-of-the-art traditional predic-
tive algorithms to validate its technological superiority in discovering strategic
deposits in complex geological provinces.

COMMITTEE MEMBERS:
Presidente - 1221251 - MARTIN ALEJANDRO MUSICANTE
Interno - 2177445 - BRUNO MOTTA DE CARVALHO
Interna - 2859606 - SILVIA MARIA DINIZ MONTEIRO MAIA
Externa ao Programa - 3325637 - POLIANA IARA DE FREITAS TOLEDO - UFRN