Felipe Santibañez-Leal

Felipe Santibañez-Leal

he/him

I&F Decision Science Assoc. Manager @ Accenture | Applied Scientific Researcher

Publication in Natural Resources Research: Ore-Waste Discrimination with Adaptive Sampling

2 minute read

Published:

Our paper “Ore-Waste Discrimination with Adaptive Sampling Strategy” has been published in Natural Resources Research. Read the paper (DOI).

From theory to practice

If the Mathematical Geosciences paper laid the theoretical foundation, this paper is where the rubber meets the road. This was always intended as the “applied companion” to the theoretical work – the proof that the formalism was not just elegant mathematics but something that could survive contact with real geological complexity. Here we take the information-driven sampling framework and apply it directly to real mining problems – specifically, ore-waste discrimination for grade control and short-term mine planning.

The central question is practical: given a limited budget of samples, how should they be distributed to best distinguish ore from waste in a mining block? The standard industry practice relies on regular grid drilling patterns. Our work demonstrates that an adaptive, information-guided strategy can do significantly better.

Validation on real cases

What makes this paper particularly satisfying is the validation. We tested the approach on three real mining cases with distinct geological settings and orebody geometries, and in each one the adaptive strategy showed clear improvement over the standard regular grid approach. The feeling of watching the algorithm place samples in geologically sensible locations – concentrating effort near ore-waste boundaries where uncertainty is highest – was remarkable. The theory was not just working in the abstract; it was making decisions that an experienced geologist would recognize as intelligent.

Sampling strategies comparison

Adaptive sampling:
select x* = argmax H(Xf | Xf ∪ {x}) — place next sample where uncertainty is highest

This is not a marginal academic improvement – it translates to better classification of ore and waste blocks, which directly impacts operational decisions and economic outcomes. It is worth emphasizing what this means in practice: mining companies overwhelmingly rely on regular drilling grids for grade control. Showing that an adaptive, information-driven approach consistently outperforms this industry standard, on real data from real mines, challenges deeply entrenched operational habits. The potential savings from better ore-waste boundary delineation are substantial.

For me, this paper closes a loop that started during my doctoral research. The theoretical framework needed to prove itself in realistic scenarios, and it did. Seeing the ideas work on actual mining data, with all the messiness and complexity that entails, was one of the most rewarding moments of my PhD journey. It confirmed that the years spent on formalization were not in vain. The code implementing the information-driven sampling framework used in both this paper and the earlier Mathematical Geosciences publication is openly available. View the code on GitHub.

You May Also Enjoy

Week in review: From developer portfolios to executive narratives

2 minute read

Published:

This week’s biggest change wasn’t technical – it was narrative. I restructured six industry portfolio entries to lead with business impact instead of algorithms. The shift reflects a real career evolution: from implementing models to owning their business outcomes.

Week in review: Vectorization, numerical stability, and biological accuracy

3 minute read

Published:

Modernizing eight research codebases at once sounds like a terrible idea, and some weeks it feels like one. But working across optics, biophysics, robotics, and geostatistics in parallel has surfaced a pattern: the same three engineering challenges keep showing up in every domain, wearing different clothes.

The first SOFI implementation in Chile, modernized

2 minute read

Published:

In 2012, I started working on Super-resolution Optical Fluctuation Imaging at SCIAN-Lab, Universidad de Chile, collaborating remotely with researchers at the III Physics Institute of the University of Gottingen in Germany. Working across continents and time zones with the Gottingen team was both challenging and exhilarating – SOFI was a relatively new technique at the time, and there was limited local expertise in Chile to draw on. Every implementation decision, from the cumulant computation pipeline to the deconvolution strategy, required careful coordination through emails and video calls, often debugging numerical issues with a six-hour time difference. The goal was to implement SOFI from the ground up – cumulant computation, synthetic quantum dot simulators, deconvolution – and apply it to biological samples. After months of work, we achieved the first successful SOFI super-resolution imaging in Chile, resolving structures below the diffraction limit using nothing more than temporal statistics from blinking fluorescent emitters. The pride of being the first group in Chile to produce these results was immense – it proved that cutting-edge computational optics could be done here, not just in European labs with larger budgets and established traditions in the field. That work contributed to a publication at SPIE Photonics West, presenting the cumulant-based framework and validation results.