When mapping a complex mathematical or physical landscape—whether it is the vibration profile of an aircraft hull, the distribution of soil liquefaction potential, or human aesthetic preferences—we face a fundamental bottleneck: evaluating the true landscape is expensive. High-fidelity finite element simulations, geotechnical site surveys, and crowdsourced human trials take significant time and resources.
The traditional approach to this problem is uniform or space-filling sampling (such as grid search or Latin Hypercube sampling). However, uniform sampling treats every region of a parameter space with equal importance. It wastes precious computational budgets precisely mapping flat, predictable plains while completely missing sharp, critical peaks or complex boundaries that fall between the grid lines.
Model-driven adaptive sampling solves this by turning data collection into an intelligent, iterative feedback loop. Instead of deciding all sample points upfront, an algorithm uses a lightweight surrogate model (or "metamodel") to evaluate the current landscape, identify the most informative regions, and choose the next sample points dynamically.
The Adaptive Feedback Loop
The core mechanism of adaptive sampling can be understood as a closed-loop control system. Rather than sampling blindly, the system continuously asks: Given what we currently know, where will the next sample provide the highest value?
┌────────────────────────────────────────────────────────┐
│ │
▼ │
[ Fit/Update Surrogate Model ] ──► [ Evaluate Acquisition Criteria ]
│
▼
[ Run High-Fidelity Evaluation ] ◄── [ Select Next Optimal Point(s) ]
By starting with a coarse initial grid, the system avoids identifying spurious patterns in sparse data, as noted by Chris Bailey‐Kellogg in his work on Spatial Aggregation for Qualitative Assessment of Scientific Computations. As the loop progresses, the surrogate model guides the sampling density, refining only in areas of ambiguity or high interest.
Depending on the goal of the characterization, researchers deploy different criteria to select the next optimal points. These criteria broadly fall into three strategies: error-guided exploration, task-driven exploitation, and anti-crowding balance.
Error-Guided Exploration
The most intuitive way to improve a surrogate model is to sample where it is most uncertain or prone to error. By targeting regions of high prediction variance or model disagreement, adaptive sampling focuses computational effort where the model's knowledge is weakest.
In high-dimensional integrated circuit design, Bill Lin 0001 et al. demonstrated in Enhanced metamodeling techniques for high-dimensional IC design estimation problems that this strategy significantly reduces the number of samples required to reach a target accuracy compared to uniform methods. Similarly, Maarten De Munck et al. utilized an adaptive response point selection strategy in A Fuzzy FRF Analysis of a Stiffened Conical Shell Structure Using an Intelligent Kriging Based Optimisation Procedure to efficiently construct Kriging surfaces for structural engineering, minimizing expensive finite element evaluations.
To systematically isolate these high-error regions, Zhao Ke et al. leveraged CV-Voronoi sequence sampling in Fast Vibration Reduction Optimization Approach for Complex Thin-Walled Shells Accelerated by Global Proper Orthogonal Decomposition Reduced-Order Model. This technique partitions the design space into Voronoi cells and uses cross-validation (CV) to find which cells exhibit the lowest model accuracy, placing new samples directly into those sensitive zones to meet global accuracy requirements with minimal data.
Task-Driven Exploitation: Extrema and Boundaries
Often, we do not need a perfectly accurate model of the entire landscape; we only care about regions that dictate critical outcomes, such as failure boundaries, optimal designs, or performance peaks.
Extrema Point Sampling
When the ultimate goal is optimization, sampling flat regions is a waste of time. Guang Pan et al. introduced a Sequential Optimization Sampling Method (SOSM) in A Sequential Optimization Sampling Method for Metamodels with Radial Basis Functions that explicitly targets the "peaks and valleys" (maxima and minima) of the current response surface. Because these extrema represent critical design conditions, refining the model at these points rapidly accelerates global optimization.
Taking this further, Bernardetta Addis et al. designed the ADORO framework in Data augmentation driven by optimization for membrane separation process synthesis. Their method uses an optimization loop to find stationary points where the surrogate model's predictions deviate from a high-fidelity simulator. It then densifies the training set by injecting perturbed data points specifically around these high-interest, high-error regions, ensuring the surrogate is highly accurate where the optimizer actually visits.
Excursion and Boundary Mapping
In safety-critical applications, the goal is often to map an "excursion set"—the region where a function exceeds a dangerous threshold. Dario Azzimonti et al. showed in Quantifying Uncertainties on Excursion Sets Under a Gaussian Random Field Prior that instead of relying on space-filling designs, intelligently selecting simulation points near the estimated boundary of the threshold drastically improves the accuracy of uncertainty quantification, saving massive computational overhead.
Preventing Collapse: The Anti-Crowding Constraint
A pure focus on error or extrema can introduce a new failure mode: sample crowding. If a surrogate model has high uncertainty or a sharp peak in one narrow, noisy region, an unconstrained adaptive algorithm might repeatedly place new samples right next to each other. This yields highly redundant information and leaves the rest of the design space completely unexplored.
To solve this, modern adaptive sampling algorithms use multi-objective acquisition functions that balance exploration and exploitation. For instance, in predicting the Tritium Breeding Ratio for fusion reactors, Petr Mánek et al. developed an algorithm in Fast regression of the tritium breeding ratio in fusion reactors that selects new data points based on a dual mandate:
- Maximize the surrogate model's error (targeting critical regions).
- Minimize a crowding distance metric (ensuring points are spatially distributed).
This geometric constraint forces the algorithm to spread its bets, maintaining global awareness while still diving deep into high-error zones. Guang Pan et al. utilized a similar balance in their SOSM framework, alternating between targeting critical extrema and under-sampled, sparse regions of the design space to guarantee robust global coverage.
Broad Applicability Across Domains
While heavily utilized in engineering simulations, the underlying intuition of adaptive sampling—allocating resource budget based on model-informed uncertainty—applies universally whenever data collection is constrained.
- Human Perception & Aesthetics: In speech perception experiments, Geoffrey Stewart Morrison demonstrated in An adaptive sampling procedure for speech perception experiments that fitting a logistic regression model to listener responses in real-time allows researchers to select the most ambiguous stimuli for subsequent trials, shortening experiments by a third. Similarly, Aoyu Wu et al. used importance- and gradient-based sampling in Learning to Automate Chart Layout Configurations Using Crowdsourced Paired Comparison to efficiently navigate massive chart configuration spaces based on crowdsourced human preferences.
- Data Streams & Sensor Networks: For smart city infrastructure, Isam Mashhour Al Jawarneh et al. showed in QoS-Aware Approximate Query Processing for Smart Cities Spatial Data Streams that sampling data based on real-time spatial characteristics rather than uniform time intervals meets strict accuracy targets with substantially less data. In civil engineering, Huajian Yang et al. designed an Adaptive Fusion Sampling Strategy Combining Geotechnical and Geophysical Data for Evaluating Two-Dimensional Soil Liquefaction Potential and Reconsolidation Settlement, which adaptively collects expensive, high-accuracy geotechnical data only until a desired level of certainty is reached alongside cheap geophysical data.
- Statistical Inference: Even in pure statistics, Dan M. Kluger showed in M-estimation under Two-Phase Multiwave Sampling with Applications to Prediction-Powered Inference that integrating machine learning predictions into multiwave sampling frameworks allows researchers to identify where cheap proxy data is most likely to be biased. This allows them to target expensive "gold-standard" measurements precisely where they will do the most work to debias the final statistical estimate.
Ultimately, adaptive sampling works because it treats information as a non-uniform commodity. By pairing a flexible surrogate model with an acquisition strategy that respects both model uncertainty and geometric spacing, it ensures that no computational or physical effort is wasted on predictable terrain.
Key Foundational Work
An adaptive sampling procedure for speech perception experiments (2006) — contributed 1 concept elements to this direction.
Spatial Aggregation for Qualitative Assessment of Scientific Computations (2004) — contributed 1 concept elements to this direction.