A New Era of Subsurface Intelligence with 4Point AI
A Note from 4Point AI Leadership
For decades, mining has been forced to make billion-dollar decisions with incomplete and biased subsurface information. Drilling is expensive, data density reflects history and budgets rather than geology, and conventional methods often smooth away the very boundaries and structures that control value and risk.
Today, we are sharing a meaningful step forward: 4Point AI's Spatial Intelligence Platform, built around our Spatially Informed Intelligence Model (SIIM). It turns sparse, biased subsurface datasets into accurate and stable predictions of mineralization trends in 3D, as well as how material is likely to behave through processing.
What makes this different is not "AI for mining" as a slogan. It is a validated intelligence layer designed for the realities of subsurface data: bias correction, geological context embedding, coherence checks, and governance-ready explainability so outputs remain physically credible under stress.
This post provides a high-level view of what's inside our Flagship Technical Report and why it matters for exploration, resource modeling, mine planning, and processing teams. The Flagship Technical Report and Executive Summary are available by request. If you want the full methodology, validation protocol, and benchmark tables, you can request access below.
Figure 32: Parity plots for cyanide-leach model predictions showing the agreement between predicted and measured values for (a) gold (Au) and (b) copper (Cu). The close clustering of points along the 1:1 line demonstrates strong predictive accuracy and low systematic bias. Under drillhole-grouped validation, the model achieves R² = 0.882 for gold and R² = 0.932 for copper.
Figure 33: Residual histograms for cyanide-leach predictions showing model error distributions for (a) gold (Au) and (b) copper (Cu). The sharp concentration around zero demonstrates minimal bias and low variance across predictions, with residuals tightly centered near zero even under strict validation conditions.
Introducing Spatial Intelligence
4Point AI rebuilds drill and geology databases into a spatially aware representation that preserves geological context, from domains and structures to mineralogical and geochemical signals, so predictions remain consistent in 3D.
Instead of learning drilling density or historical sampling patterns, the platform is designed to learn transferable geological relationships. Performance is validated explicitly in new holes, sparse areas, boundary conditions, and unfamiliar regimes. Clients are not buying a one-off model. They gain an adaptable Spatial Intelligence engine that can be retargeted as priorities evolve.
Performance Where It Actually Matters: Under Stress
Mining teams do not need models that look good on easy splits. They need models that remain reliable when geology changes, sampling thins out, and downside risk increases.
Exploration-Scale Prediction That Stays Unbiased and Stable
In regional-scale mineral prediction, error remains centered near zero (mean deviation around 0.24), with residuals tightly concentrated and strong variance capture (approximately 88% to 90%). The parity and residual plots above illustrate this stability directly, showing tight fit and minimal systematic bias under strict validation.
In separate tests focused on directional behavior, orientation errors concentrate within plus or minus five degrees, supporting reliable trend definition rather than diffuse or noisy spatial outputs.
Generalization to Unseen Drillholes and Out-of-Range Conditions
Generalization is tested directly, including strict external validation. For clay prediction, internal R-squared exceeds 0.97 and external performance holds at R-squared = 0.875 under out-of-range feature combinations.
For cyanide-leach prediction, drillhole-grouped validation (entire holes withheld) achieves R-squared = 0.882 for gold and R-squared = 0.932 for copper, with residuals remaining centered near zero.
Figure 27: R² heatmap for strict clay prediction under fault-uniform validation, showing model performance across 15 clay mineral types and four cumulative probability thresholds (10%, 25%, 50%, 100%). Values consistently exceed 0.94 across all conditions, with several clay types (kaolinitewx, muscovitecillite, serpentine) achieving perfect prediction (R² = 1.00). This fault-aware validation partitions training and test data along geological structures, demonstrating that the model generalizes to geologically distinct regions rather than simply interpolating between nearby samples.
Table 6: Model performance for cyanide-leach prediction under different validation schemes.
| Target | R² (OOF) | RMSE (OOF) | R² (GroupCV) | RMSE (GroupCV) |
|---|---|---|---|---|
| Au (cyanide-leach) | 0.831 | 0.980 | 0.882 | 0.797 |
| Cu (cyanide-leach) | 0.934 | 0.139 | 0.932 | 0.138 |
Table 6: Model performance for cyanide-leach prediction under different validation schemes. OOF (Out-of-Fold) represents standard cross-validation, while GroupCV withholds entire drillholes to simulate prediction in unsampled areas. The model maintains strong performance under both schemes, with R² values of 0.882 for Au and 0.932 for Cu under the stricter drillhole-grouped validation.
Benchmarking vs Classical Methods: Not Marginal Gains
In a controlled comparison against Kriging, Co-Kriging, and CatBoost across three scenario regimes and fifteen geological conditions, SIIM consistently reduces prediction error by 50% to over 70%. In sparse and boundary settings, reductions reach 60% to 90%, and extreme outliers are reduced by 50% to 80%.
Even in low-complexity domains where Kriging approaches its theoretical optimum, SIIM maintains an advantage, typically improving performance by 20% to 40%.
In practice, this is the difference between confident step-outs and expensive infill, between stable schedules and surprise variability.
| Domain / Target | Scenario | Description |
|---|---|---|
| Enriched Cu – Au | HC | High-complexity condition |
| Enriched Cu – Au | LC | Low-complexity condition |
| Enriched Cu – Au | HK | High-continuity condition |
| Enriched Cu – Au | 15C | Fifteen-condition benchmark |
Experimental scenarios used in the controlled benchmarking comparison. The Enriched Cu Domain was selected for its combination of strong conditional continuity and pronounced grade variability, providing an appropriate stress-test environment for comparing SIIM with classical geostatistical and machine-learning methods. Three case-level complexity scenarios (HC: High Complexity, LC: Low Complexity, HK: High Continuity) and fifteen geological conditions were evaluated for both Au and Cu targets.
Figure 37: Predictive model error comparison for Au (ppm) in the Supergene Enrichment Zone. The top row shows case-based performance across three complexity scenarios (HK, HC, LC), while the bottom row shows condition-based performance across fifteen geological conditions including lithology transitions, fault zones, sparse data regions, and nonlinear feature interactions. SIIM (green) consistently achieves the lowest mean absolute error, reducing prediction error by 80–97% relative to Kriging and Co-Kriging and 80–86% relative to CatBoost. Even in low-complexity settings where Kriging operates near its theoretical optimum, SIIM maintains a measurable performance advantage.
From Grade to Behavior: Predicting What the Rock Will Do
A core advantage of the 4Point AI Spatial Intelligence Platform is that it is not limited to a fixed set of outputs. If a property varies spatially and materially impacts decisions, SIIM can learn it, quantify it, and map it in 3D. The platform unifies geology and metallurgy so predictions remain internally consistent across the orebody and through processing.
Below are two examples from the Flagship Technical Report. They are examples, not limits.
| Example 1: Clay Prediction as an Operational Risk Map
Clay can be a leading indicator of permeability, recovery, reagent consumption, and material handling risk. The flagship results show clay probability can be mapped with high accuracy and stable performance under fault-aware and strict external tests. This enables earlier identification of problematic material and more targeted metallurgical programs, rather than reacting after issues emerge in the plant.
↑ View Figure 27: Clay Prediction Results
| Example 2: Predicting Cyanide-Leachable Metals
Cyanide-leachable gold and copper represent the fraction of metal actually recoverable under standard leaching. SIIM distinguishes oxidized versus sulfide domains and captures copper responses consistent with weathered oxides and carbonates versus fresh sulfides. This supports continuous spatial planning of recoverability instead of relying on sparse composites or domain-wide assumptions.
↑ View Figure 32: Cyanide-Leach Prediction Results
Coherence Checks That Build Trust
Beyond accuracy, the platform applies coherence and interpretability checks so predictions remain physically credible and governance-ready. In the flagship results, high clay probability zones coincide spatially with reduced gold and copper recoveries, matching known process limitations and reinforcing confidence in deployment.
The same approach extends naturally to other spatial targets, including alteration intensity, hardness proxies, deleterious elements, density, domain boundaries, and uncertainty-aware ore control indicators.
Built for the Real Constraints of Mining Data
The flagship report is grounded in real-world datasets across multiple sites and scales, selected to test consistency under varying mineralization styles, lithological complexity, and sampling density.
Data are cleaned, standardized, and organized into a relational structure that preserves spatial and geological relationships so validation remains traceable and reproducible.
Mining datasets are incomplete, inconsistent, and multimodal by nature. This platform is designed to operate within those constraints, capturing more realistic and physically consistent patterns than distance-only interpolation by integrating geological context alongside geometry.
1
Raw Mining Data
Sparse, biased, multimodal
Click to explore
2
Data Engineering
Clean, standardize, trace
Click to explore
3
Spatially Aware Representation
Geological context preserved
Click to explore
4
Bias-Corrected Sampling
Learn geology, not drill patterns
Click to explore
5
Spatial Intelligence Engine
SIIM
Click to explore
6
Spatial Intelligence Outputs
Decision-ready predictions
Click to explore
New drilling and assays continuously update training and validation
Trust + Governance Layer
Built-in
Stress-Tested Validation
External tests, drillhole-grouped validation, condition-based stress testing under geological complexity
Explainability + Coherence
Outputs are physically credible, auditable, and governance-ready
Cross-Signal Consistency
Example: clay behavior aligns with recovery response, where signals constrain each other
Select any stage above to explore details
What This Unlocks Across the Mine Lifecycle
The practical impact is not abstract model metrics. It has fewer surprises, sharper decisions, and reduced volatility.
Exploration: improved targeting and step-out confidence, higher return per meter drilled, faster conversion from discovery to definition.
Resource modeling and planning: sharper ore-waste boundaries, reduced dilution and ore loss, fewer redesigns, more predictable reconciliation.
Processing and geometallurgy: targeted testwork, proactive reagent planning, better blending and scheduling to protect throughput and recovery.
Risk and investment: auditable validation, stability across conditions, and interpretable checks that strengthen confidence with boards, partners, and financiers.
Why 4Point AI Is Different
4Point AI is designed around subsurface reality:
- Outperforms distance-only interpolation by embedding geological context and structural conditioning, not just coordinates.
- Unifies geology and metallurgy so predictions remain internally consistent.
- Demonstrates generalization through strict external tests, drillhole-grouped validation, and condition-based stress testing.
- Provides explainability and transparency aligned with physical understanding and governance needs.
- Delivers reusable platform capability rather than one-off models that must be rebuilt each time.
What's Next
The Flagship Technical Report shows that Spatial Intelligence delivers a substantial reduction in prediction error across diverse geological regimes, including the settings where conventional approaches are most fragile.
The result is a deployable framework that supports better drilling decisions, more reliable planning inputs, and more predictable recovery and cost performance.
If you want to see what this looks like on your asset, the fastest path is simple: bring your existing drillhole and domain data, and we will show you how Spatial Intelligence changes targeting confidence, boundary risk, and processing predictability.
Dig into the future
of mining
Leaders in Geophysical Artificial Intelligence & Machine Learning
About the authors