⚡ electrical engineering

Algorithmic Blast Design through Deep Reinforcement Learning: A Transferable Simulation Framework from Central Asian Copper Megapits to Peruvian Mining and Public Spaces

This paper proposes and evaluates a Deep Reinforcement Learning framework based on Proximal Policy Optimization that, through simulation, demonstrates significant improvements in fragmentation uniformity and vibration control for Kazakh copper megapits while outlining the technical and regulatory conditions required for its transferability to Peruvian mining operations and urban public spaces.

Original authors: PAUL RICARDO PRUDENCIO GALVEZ

Published 2026-07-16

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: PAUL RICARDO PRUDENCIO GALVEZ

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to break a giant, stubborn chocolate bar into perfect, bite-sized pieces. If you just smash it with a hammer, some pieces will be huge chunks, others will turn to dust, and the whole thing might crack in a way you didn't expect. This is exactly what happens in massive copper mines. They need to blast rock into just the right size so trucks can carry it and machines can crush it. For decades, miners have used old-school "rule-of-thumb" formulas to decide where to put their explosives. But rocks aren't like chocolate; they are messy, uneven, and full of surprises. These old rules often miss the mark, causing too much shaking (which can hurt nearby towns) or leaving rocks too big to handle.

Enter a new kind of problem-solver: an Artificial Intelligence (AI) that learns by playing a game. Instead of following a fixed rulebook, this AI is like a video game character that tries thousands of different ways to smash the rock. Every time it tries a new pattern, it gets a "score" based on how well the rock broke and how little the ground shook. Over time, the AI learns the secret moves to get the perfect break every time. This isn't just about saving money; it's about being safer and smarter. If we can teach machines to blast better, we can mine copper more efficiently and keep the ground from shaking too hard, even when we are digging near busy cities.

The Paper's Big Idea: Teaching Rock-Blasting Robots to Play

This paper is about a scientist named Paul who built a virtual video game to teach an AI how to blast copper mines better than the old methods. He didn't go out to a real mine with a truckload of explosives. Instead, he created a "digital twin"—a super-smart computer simulation—where he could test his ideas without any real-world danger.

The Game Setup
Paul set up a virtual world representing a giant copper mine in Kazakhstan (a place with huge copper deposits). In this world, he programmed an AI agent using a method called "Deep Reinforcement Learning" (specifically, an algorithm called PPO). Think of this AI as a digital drill sergeant. Its job is to decide three things for every blast:

How far apart to drill the holes (burden and spacing).
How much explosive to put in each hole.
The exact timing of the explosion delays (so the rocks hit each other at the perfect moment).

The AI's goal was simple: break the rocks into the most uniform size possible while keeping the ground vibration as low as possible. It played this game 10,000 times for every block of rock, learning from every mistake and success.

The Results: A Virtual Victory
When Paul compared the AI's performance to the traditional, old-school "Kuz-Ram" method (the standard rule-of-thumb used for years), the AI showed some impressive improvements in the simulation:

Better Rock Sizes: The AI reduced the messiness of the broken rocks by 18% to 34%. This means the rocks were much more uniform, which is a dream for mining trucks and crushers.
Less Shaking: The AI lowered the peak ground vibration (measured as Peak Particle Velocity, or PPV) by 15% to 30%. This is huge because less shaking means less risk to nearby buildings and communities.
Less Waste: In the trickiest, most uneven rock scenarios, the AI used 8% to 19% less explosive to get the job done.

The Catch: It's Still a Simulation
Here is the most important part: Paul is very honest about what this means. These numbers are not from a real mine. They are from a computer game built using data from 14 different scientific studies. The paper explicitly states that these results are "methodologically plausible" but not proven in the real world yet. The AI hasn't actually blasted a real mountain in Peru or Kazakhstan. The results are consistent with what other scientists have seen in similar studies, but they are just estimates from a virtual environment.

Can We Bring This to Peru?
The paper then asks: "Could this work in Peru's famous copper mines?" Paul built a comparison chart to check.

Good Candidates: The mines at Cerro Verde and Quellaveco look very similar to the virtual mine in the simulation (they are both "copper porphyry" deposits). The paper suggests these places would be the best spots to try this technology first.
Tricky Candidates: The Antamina mine has a different type of rock (polymetallic skarn) that is more chaotic, so the AI would need a lot of extra training to work there. Las Bambas has complex cracks in the rock, which might actually be a good place for an AI because it can adapt quickly, but it would need more sensors.

From Mines to City Streets
The paper takes a wild leap at the end: What if we use this same AI to blast in cities? Imagine digging a new subway line in Lima or stabilizing a hillside near a school. In a city, you can't have the ground shaking like a mine. The paper suggests that if we adapted the AI's "reward system" to care more about keeping the city quiet than making money, it could help manage vibrations in urban areas. However, this would require much stricter rules and a lot more community communication.

What's Needed to Make It Real?
Before any of this happens in the real world, the paper lists the "shopping list" for a pilot test:

Machinery: Drills with GPS, super-precise electronic detonators (that can time explosions to the millisecond), and sensors to measure vibration and rock size.
People: A team that mixes mining engineers with data scientists and geologists.
Safety: A strict plan to make sure no one gets hurt and the community is happy.

The Bottom Line
This paper proposes a brilliant, futuristic way to blast mines using AI. It shows that in a computer simulation, this AI can break rocks better and shake the ground less than our current methods. But the author is careful to say: "We haven't tested this on a real mountain yet." It's a very promising map for a journey, but the journey itself—testing it in a real Peruvian mine or a city street—hasn't started. The next step is to take this digital idea, build the real equipment, and try it out in a controlled pilot project to see if the magic works outside the computer.

Technical Summary: Algorithmic Blast Design through Deep Reinforcement Learning

Problem Statement
Current bench blasting practices in porphyry copper megapits, such as the Aktogay and Bozshakol complexes in Kazakhstan, predominantly rely on deterministic empirical models (e.g., Kuz-Ram). These models fail to adequately capture the spatial variability of rock mass properties, leading to operational inefficiencies including over- or under-fragmentation, excessive ground vibration, and increased downstream energy consumption. While Deep Reinforcement Learning (DRL) offers a promising adaptive alternative for sequential optimization, its specific application to bench-scale blast scheduling remains underexplored. Furthermore, there is a lack of systematic assessment regarding the transferability of such algorithmic frameworks from Central Asian contexts to the distinct geological, regulatory, and social environments of Peruvian mining operations and urban public spaces.

Methodology
The study employs a computational simulation and comparative-documentary research design, avoiding intervention in active mining operations or the use of proprietary data. The core methodology consists of four integrated components:

Simulation Environment: A simplified digital twin of the rock mass was constructed using the Discrete Element Method (DEM). This environment was parameterized with statistical distributions of Uniaxial Compressive Strength (UCS), Rock Quality Designation (RQD), burden, spacing, and delay sequences. These parameters were calibrated exclusively using ranges reported in fourteen peer-reviewed studies (2019–2026) on blast-induced fragmentation and vibration.
DRL Agent Formulation: The blast design problem was modeled as a Markov Decision Process (MDP). A Proximal Policy Optimization (PPO) agent with an Actor-Critic architecture was trained over 10,000 simulated episodes per block.
- State Space ( $S$ ): Included rock properties (UCS, RQD, density), current design parameters (burden, spacing, depth), and historical fragmentation performance.
- Action Space ( $A$ ): Continuous adjustments to burden, spacing, and specific explosive charge, alongside discrete selection of inter-hole delay sequences.
- Reward Function ( $R$ ): A weighted objective minimizing fragmentation standard deviation, peak particle velocity (PPV), and specific explosive consumption, while maximizing simulated loading-hauling productivity.
Comparative Analysis: The PPO agent's performance was benchmarked against a deterministic Kuz-Ram-type baseline under identical simulation assumptions.
Transferability Assessment: A qualitative-quantitative matrix was constructed to compare the Central Asian reference context with four major Peruvian copper operations (Antamina, Cerro Verde, Quellaveco, Las Bambas) and urban public space projects. This analysis utilized only public production data from Peru's Ministry of Energy and Mines (2024) and publicly available geological characterizations.

Key Results
Within the literature-calibrated simulation environment, the DRL framework demonstrated the following performance improvements relative to the deterministic baseline:

Fragmentation Precision: A reduction in the standard deviation of fragmentation by 18–34%.
Vibration Control: A reduction in Peak Particle Velocity (PPV) by 15–30%.
Explosive Efficiency: A reduction in specific explosive consumption of 8–19% in scenarios with high geological heterogeneity.

These magnitudes are consistent with improvements reported in independent machine-learning-based optimization studies (e.g., PPV reductions of 12–40% in reviewed literature), supporting the methodological plausibility of the framework.

Transferability Findings
The study identified varying levels of compatibility for deploying this framework in Peru:

High Compatibility: Cerro Verde and Quellaveco were identified as the most suitable candidates due to their relatively homogeneous copper porphyry geology, which structurally resembles the Central Asian reference deposits.
Medium/Lower Compatibility: Antamina (polymetallic skarn) presents greater lithological heterogeneity, requiring substantial agent retraining. Las Bambas (complex faulting) represents an intermediate case where DRL's adaptive capacity could add value but demands significant instrumentation investment.
Urban Application: Applying precision blasting to Metropolitan Lima's public spaces would require vibration thresholds an order of magnitude stricter than remote mining, necessitating a reweighting of the reward function to prioritize social variables over productivity.

Significance and Claims
The paper positions this work as a methodologically plausible pathway toward precision blasting, distinguished by three primary contributions:

Operational Focus: It explicitly integrates bench-level blast scheduling into a reinforcement-learning framework, moving beyond strategic planning or fleet dispatching.
Transparency and Integrity: By utilizing a literature-calibrated simulation rather than unverifiable proprietary data, the study establishes a reproducible baseline for comparing DRL against deterministic models.
Cross-Domain Extension: It proposes a conceptual extension of mining RL frameworks to urban infrastructure vibration management, a domain rarely connected to mining literature.

Limitations and Modesty of Claims
The authors explicitly state that the quantitative results are estimates from a simulation environment, not field measurements. Consequently, the study does not claim definitive operational performance. The conclusions are framed as "methodologically plausible" and consistent with existing literature, emphasizing that:

Field validation on a deposit-by-deposit basis is an indispensable prerequisite for implementation.
Successful deployment in Peru requires specific adaptations in machinery (e.g., GPS/MWD drills, electronic detonators), materials, and personnel competencies (blasting engineers with data science skills).
Regulatory adaptation and community consultation are critical, particularly for urban applications.

The study concludes that while the DRL framework offers a robust theoretical alternative to empirical models, its real-world benefits must be validated through controlled pilot trials before operational adoption.

Technical Summary: Algorithmic Blast Design through Deep Reinforcement Learning

More like this