hi there, I'm
Vlad Alekseev
Geophysics PhD candidate at UNSW Sydney and lead computer vision engineer. I build deep learning for 3D data, from seismic and rock micro-CT to CT and CBCT, with 6+ years in production.
Survey-conditioned pre-training for label-efficient seismic fault detection
Timeline
PhD research: deep learning for seismic interpretation
Faults: three label-free ways to pre-train on the unlabelled target survey, then fine-tuning on a few labelled sections. Scored against an expert and eight interpreters on F3 CRACKS.
Trained on one interpreter's picks, the model beats all eight interpreters: line F1 0.58 against 0.47. With no labels, a survey-calibrated generator beats five of the eight.
Relative geologic time: a systematic evaluation of a published 3D HRNet RGT model, trained on 2,000 synthetic volumes, on held-out synthetics and on F3 and Thebe field data, with sparse horizon guidance.
Horizon guidance cuts the error from 4.05 to 0.74 samples on held-out synthetics. A memory-lean rewrite runs a full volume in 2.2 s on one 16 GB GPU.
UNSW Digital Twin: machine learning on real campus data
A curated offline dataset of the UNSW Kensington campus (aerial imagery, buildings, occupancy, CPTU soundings, weather, electricity) and unsw_twin, the Python toolkit behind the tutorials for CVEN4801, a machine learning course in civil engineering.
Digital rock from core micro-CT
3D rock models from micro-CT and SEM of core samples to estimate permeability in tight, microporous rock where lab measurement is not feasible. Denoising, multi-mineral segmentation, multi-resolution registration. Led the team from 2022.
10× faster image registration
3D CBCT AI and DentalMesh Studio
ML: end-to-end CBCT pipeline with segmentation of anatomy, airway and teeth, FDI tooth numbering, cephalometric landmarks, clinical reformats and PDF reports. Synthetic CBCT with metal artefacts for robustness.
Studio: a 3D CBCT viewer where every action is also a typed command in English or Russian, with tissue isolation, measurements and full undo.
In commercial use: labs and clinics buy segmentation, reformat and cephalometry reports per CBCT scan.
ORION: computer vision for school safety
ORION turns existing school cameras into an always-on safety monitor, with behaviour analysis, stranger and restricted-zone alerts, and an interactive campus map.
Product design for AI tools in oil and gas
Designed and prototyped three AI web apps and managed their front-end development: DeepCore for automated core description, AIDrilling for drilling incident prediction and 4FIELD for fast field assessment. Also led the company rebrand.
Survey-conditioned pre-training for label-efficient seismic fault detection
V. Alekseev, P. Makuluni, R. T. Armstrong, N. Khalili, S. R. Clark. Extended abstract and short oral talk, session 3.5.
Trained on one interpreter's picks, the model beats all eight interpreters: line F1 0.58 against 0.47. With no labels, a survey-calibrated generator beats five of the eight.
Papers on SEM, micro-CT and core images
- Multi-mineral segmentation of SEM images using deep learning techniquesV. Alekseev, D. Orlov, D. Koroteev · SPE Russian Petroleum Technology Conference
- Mineral mapping of rock micro-images using computer vision and deep learning methodsV. Alekseev, D. Orlov, D. Koroteev · ProGREss'21
- An intellectual system for fast and reliable automated core description developmentE. Baraboshkin, A. Demidov, D. Orlov, V. Alekseev, et al. · 82nd EAGE Annual Conference
- Automated full-bore core description application for production purposesE. Baraboshkin, A. Demidov, et al., V. Alekseev · Data Science in Oil and Gas 2021