Space-Filling One-Factor-At-A-Time Designs

arXiv:2606.02533 · stat.ME · Submitted 2026-06-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Space-Filling One-Factor-At-A-Time Designs".

Jane: Here's the gist of the paper. Space-filling designs are proposed as a new class of screening designs that improves space-fillingness while retaining their screening capability.

Tom: This is important because traditional space-filling designs cannot perform factor screening without modeling assumptions, which makes them inefficient when only a small subset of input factors is important.

Jane: Space-filling designs are widely used in deterministic computer experiments to place points everywhere in the input space with as few gaps as possible, leading to robust performance for learning an unknown response surface.

Tom: Factor screening is known as “factor screening” in sensitivity analysis and “variable selection” in statistics, and it can be done by changing one factor at a time which is the simplest way to perform factor screening without fitting a surrogate model or imposing additional assumptions.

Title and authors: Tom: Let’s talk about who wrote this. The paper is titled "Space-Filling One-Factor-At-A-Time Designs". It’s authored by Wei Yang Yu and V Roshan Joseph from Georgia Institute of Technology, Atlanta, GA.

Jane: They have a solid background in experimental design and sensitivity analysis, which makes sense given what they're trying to do here.

Lu: Lu knows this area well; the authors are building on existing work about space-filling designs and factor screening methods like MOFAT designs.

Meng: I wonder how much of their proposed method actually translates to a real deployment scenario, you know, something that doesn't require super complex computation just to run the design itself.

Lalam: It seems like the focus is on creating a design structure that inherently balances those two goals—space-fillingness and screening capability—which is what they call SOFT designs.

The paper's summary: Tom: So, their core idea is this new class of designs called Space-Filling One-Factor-At-A-Time Designs. They aim to improve the space coverage while keeping that one factor at a time screening ability intact.

Jane: They point out that standard space-filling designs, even though they cover the input space well, can't do factor screening without adding assumptions or fitting a surrogate model first, which is where they get inefficient when only a small set of inputs matters.

Lu: The paper shows that their proposed SOFT designs achieve optimal space-filling properties while still retaining that screening ability because of its one-factor-at-a-time structure.

Meng: That makes sense, but I'm wondering if this improvement is just theoretical coverage, or if it actually helps when you're dealing with high dimensional data where those gaps are really big.

Lalam: They show that these SOFT designs offer clear advantages over existing designs in numerical examples, and they even compare their screening performance to MOFAT designs, suggesting the space-filling improvement doesn't hurt the factor screening part.

The paper's improvements: Tom: So what are they actually doing differently? They introduce SOFT designs which have this structure that gives them great coverage and good screening ability at the same time.

Jane: And then they developed Augmented SOFT, or A-SOFT designs, where they allocate some of the experimental budget to points that aren't restricted by the one-factor-at-a-time rule.

Lu: The goal in constructing these designs is really about finding a sweet spot—optimally choosing the factor levels, the specific one factor at a time structure, and then figuring out how to set up those base runs to make it as space-filling as possible.

Meng: Optimizing three elements at once sounds complicated. What’s the actual math driving this optimization? Is it just minimizing some kind of error metric?

Lalam: They use minimizing the mean squared error under a Gaussian process surrogate model, which they say is equivalent to maximizing a quadratic form called Q of D with alpha and theta.

Conclusion: Tom: So to wrap up, these A-SOFT designs are quite flexible because you can split the budget—about half for the SOFT part and the rest for augmentation. It seems like a robust way to balance screening and space-fillingness.

Jane: Exactly. They show that this balancing act is important because both space-fillingness and factor screening are crucial when you need good predictive performance in complex systems, especially in high dimensions.

Lu: The implication is that for experiments where effect sparsity is common, we can get better coverage without sacrificing the ability to pinpoint which factors are truly influential.

Meng: If this holds up under real traffic conditions, it means we could potentially design experiments that are more efficient at finding important inputs without having to build a massive model just to start.

Lalam: The Space-Filling One-Factor-At-a-Time Designs paper shows that by structuring the experiment this way, we get both good coverage and good screening simultaneously, which is a solid foundation for future experimental setups.

H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology

stat.ME

Submitted: 2026-06-01

Updated: 2026-10-06

Code: https://github.com/weiyangyu/SOFT

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 87/100

The gist: Space-filling designs are proposed as a new class of screening designs that improves space-fillingness while retaining their screening capability.

Key concepts

Space-filling Designs
These are experimental designs used to place points throughout the entire input space with minimal gaps. They are excellent for learning unknown response surfaces because they ensure good coverage across all input variables, making them robust for prediction.
Factor Screening
This is a method used in sensitivity analysis to identify which input factors have the most significant influence on the output. It is typically done by changing one factor at a time to see how much the output changes, which helps select only the most important variables.
SOFT Designs
These are a new class of One-Factor-at-a-Time (OFAT) designs that achieve optimal space coverage while keeping their screening capability. They are structured specifically to provide excellent input space exploration alongside the necessary one-factor comparisons for screening.
A-SOFT Designs
These designs combine SOFT properties with augmentation, allowing some budget to be used on points not restricted by the OFAT structure. This creates a flexible tradeoff between achieving good factor screening and maximizing space-filling coverage.

Terminology

Summary

Space-filling designs are proposed as a new class of screening designs that improves space-fillingness while retaining their screening capability. p. 0

Where the paper says it

Recently developed screening designs, such as MOFAT designs, can identify important factors without a modeling step but lack space-filling properties, limiting their usefulness for surrogate modeling. In this article, we propose a new class of screening designs that improves the space-fillingness while retaining their screening capability.

This is important because traditional space-filling designs cannot perform factor screening without modeling assumptions, which makes them inefficient when only a small subset of input factors is important. p. 0
Where the paper says it

Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology, Atlanta, GA 30332 October 7, 2026 Abstract Space-filling designs are commonly used in deterministic computer experiments. However, they cannot do factor screening without any modeling assumptions, which makes them inefficient when only a small subset of input factors is important.

Background on Existing Designs

Space-filling designs are widely used in deterministic computer experiments to place points everywhere in the input space with as few gaps as possible, leading to robust performance for learning an unknown response surface. Factor screening is known as “factor screening” in sensitivity analysis and “variable selection” in statistics, and it can be done by changing one factor at a time which is the simplest way to perform factor screening without fitting a surrogate model or imposing additional assumptions. p. 2

Where the paper says it

In contrast, changing one factor at a time is the simplest and most intuitive way to perform factor screening without fitting a surrogate model or imposing additional assumptions. However, a single one-factor-at-a-time (OFAT) design can only estimate local sensitivity and fails to explore the entire input space or detect interactions (Czitrom, 1999; Saltelli and Annoni, 2010).

However, a single one-factor-at-a-time (OFAT) design can only estimate local sensitivity and fails to explore the entire input space or detect interactions. p. 2
Where the paper says it

In contrast, changing one factor at a time is the simplest and most intuitive way to perform factor screening without fitting a surrogate model or imposing additional assumptions. However, a single one-factor-at-a-time (OFAT) design can only estimate local sensitivity and fails to explore the entire input space or detect interactions (Czitrom, 1999; Saltelli and Annoni, 2010).

Development of Screening Designs

The paper introduces a new class of OFAT designs called Space-filling One-Factor-at-a-Time (SOFT) designs that achieve optimal space-filling properties while retaining their screening capability. These SOFT designs are shown in Figure 1(d), which provides not only excellent coverage of the input space but also good screening ability due to its OFAT structure. p. 3

Where the paper says it

The SOFT design is shown in Figure 1(d). We can see that it provides not only excellent coverage of the input space but also good screening ability due to its OFAT structure.

Furthermore, Augmented SOFT (A-SOFT) designs are developed to allocate part of the budget to spacefilling points that are unrestricted by the OFAT structure, providing a tradeoff between factor screening and space-fillingness. p. 4
Where the paper says it

To alleviate this limitation, we further develop Augmented SOFT (A-SOFT) designs, which allocate part of the budget to spacefilling points that are unrestricted by the OFAT structure. This improves the flexibility for exploring the input space at the cost of fewer OFAT comparisons and, consequently, some loss in factor-screening accuracy.

Optimization Criteria for Design Construction

The goal is to optimally choose three elements—factor levels, OFAT structure, and the base runs—so that the design becomes as space-filling as possible. The space-filling criterion used is minimizing the MMSE under a GP surrogate, which is equivalent to maximizing the quadratic form Q(D; α, θ). p. 1

Where the paper says it

Surrogate models, such as Gaussian processes (GPs, Gramacy (2020)), are then employed to approximate the expensive computer model, making optimization and exploration more efficient in future investigations. Space-filling designs are widely used as experimental designs in deterministic computer experiments (Joseph, 2026, Ch.

To achieve optimal one-dimensional projections, the levels in L must be distinct and equally spaced with a perturbation magnitude of ∆. p. 8
Where the paper says it

equally spaced levels given by {0. To maximize the capacity of identifying important factors, A and B are chosen to maximize the expected total Sobol’ indices under the assumption that the responses are generated from a Brownian process (Zhang and Apley, 2014).

The optimal pairing matches each level with the one located lδ distance away, leading to a transformation T∗l(x) = x ± lδ.

Performance Evaluation and Findings

Numerical studies demonstrate that SOFT designs achieve strong space-filling performance across different criteria and outperform MOFAT designs, especially in high-dimensional settings. For factor screening, the screening performance of SOFT is comparable to that of MOFAT, showing that the improved space-fillingness does not come at the expense of its factor-screening capability. p. 4

Where the paper says it

This improves the flexibility for exploring the input space at the cost of fewer OFAT comparisons and, consequently, some loss in factor-screening accuracy. The augmented design therefore provides a tradeoff between factor screening and space-fillingness, with the goal of achieving more robust surrogate-modeling performance.

Finally, A-SOFT designs provide a more robust performance over other competing designs because they seem to nicely balance the two objectives of screening and space-fillingness. p. 4
Where the paper says it

This improves the flexibility for exploring the input space at the cost of fewer OFAT comparisons and, consequently, some loss in factor-screening accuracy. The augmented design therefore provides a tradeoff between factor screening and space-fillingness, with the goal of achieving more robust surrogate-modeling performance.

Conclusion on Robustness

A-SOFT designs are very flexible because they allow assigning about half of the runs to the SOFT component and the remaining to augmentation, which makes A-SOFT designs very flexible. Moreover, A-SOFT designs provide a more robust performance over other competing designs because they seem to nicely balance the two objectives of screening and space-fillingness, both of which are important for prediction. p. 4

Where the paper says it

This improves the flexibility for exploring the input space at the cost of fewer OFAT comparisons and, consequently, some loss in factor-screening accuracy. The augmented design therefore provides a tradeoff between factor screening and space-fillingness, with the goal of achieving more robust surrogate-modeling performance.

The trade-off between space-fillingness and factor-screening capability is illustrated in Figure 12. p. 3
Where the paper says it

Therefore, in this article, we propose a new class of OFAT designs, called Space-filling One-Factor-at-a-Time (SOFT) designs, that achieves optimal space-filling properties while retaining their screening capability. The SOFT design is shown in Figure 1(d).

Supplementary Materials

Proof of Propositions, graphical illustrations, details of test functions, and additional simulations are included in the supplementary file. All scripts required to reproduce all the simulation results are available at https://github. p. 4

Where the paper says it

Several numerical studies are provided in Section 4 to demonstrate the advantages of the proposed method. Motivated by these results, Section 5 introduces A-SOFT designs.

com/weiyangyu/SOFT. p. 3
Where the paper says it

Figure 1: Illustration of four OFAT-type designs: (a) a realization of Morris design, (b) optimal Morris design using the method of Campolongo et al. (2007), (c) MOFAT design, and (d) the proposed SOFT design.

Acknowledgments

This research is supported by a U.S. National Science Foundation grant DMS-2310637.

References

The paper cites numerous works related to experimental design, Gaussian process modeling, and sensitivity analysis throughout the text and also lists specific references for various methods like MmLHD, MOFAT, and Sobol’.

Key References Cited

Ba, S., SLHD: Maximin-Distance (Sliced) Latin Hypercube Designs. p. 1

Where the paper says it

Since effect sparsity is common in almost all systems— meaning that only a small subset of input factors is influential to the output—considerable effort has been devoted to improving the projections of space-filling designs onto lowerdimensional subspaces. To improve the one-dimensional projections, Morris and Mitchell (1995) proposed using a maximin Latin hypercube design (MmLHD) that maximizes the minimum distance among the points in a Latin hypercube design (LHD).

R package version 2. p. 3
Where the paper says it

mon OFAT structure, Xiao et al. (2023) proposed an improved version called the maximum one-factor-at-a-time (MOFAT) design.

1-1.

Xiao, Q., Joseph, V. p. 0

Where the paper says it

Space-Filling One-Factor-At-A-Time Designs Wei-Yang Yu and V. Roshan Joseph∗ H.

p. 0
Where the paper says it

Space-Filling One-Factor-At-A-Time Designs Wei-Yang Yu and V. Roshan Joseph∗ H.

p. 0
Where the paper says it

Space-Filling One-Factor-At-A-Time Designs Wei-Yang Yu and V. Roshan Joseph∗ H.

p. 0
Where the paper says it

Space-Filling One-Factor-At-A-Time Designs Wei-Yang Yu and V. Roshan Joseph∗ H.

p. 0
Where the paper says it

Space-Filling One-Factor-At-A-Time Designs Wei-Yang Yu and V. Roshan Joseph∗ H.

p. 0
Where the paper says it

Space-Filling One-Factor-At-A-Time Designs Wei-Yang Yu and V. Roshan Joseph∗ H.

R., and Ray, D. M., MOFAT: Maximum One-Factor-at-a-Time Designs. p. 7
Where the paper says it

(2023) proposed an improvement of the Sobol’ design, which is called the maximum one-factor-at-a-time (MOFAT) design. In their construction, A and B are two Latin hypercube designs (LHDs, McKay et al.

R package version 1. p. 3
Where the paper says it

mon OFAT structure, Xiao et al. (2023) proposed an improved version called the maximum one-factor-at-a-time (MOFAT) design.

0.

Song, D., Joseph, V. R., Efficient active learning strategies for computer experiments. p. 1

Where the paper says it

Space-filling designs are widely used as experimental designs in deterministic computer experiments (Joseph, 2026, Ch. These designs aim at placing points everywhere in the input space with as few gaps as possible, leading to robust performance for learning an unknown response surface.

p. 1
Where the paper says it

Space-filling designs are widely used as experimental designs in deterministic computer experiments (Joseph, 2026, Ch. These designs aim at placing points everywhere in the input space with as few gaps as possible, leading to robust performance for learning an unknown response surface.

Johnson, M. E., Moore, L. M., and Ylvisaker, D., Minimax and maximin distance designs. p. 1

Where the paper says it

Since effect sparsity is common in almost all systems— meaning that only a small subset of input factors is influential to the output—considerable effort has been devoted to improving the projections of space-filling designs onto lowerdimensional subspaces. To improve the one-dimensional projections, Morris and Mitchell (1995) proposed using a maximin Latin hypercube design (MmLHD) that maximizes the minimum distance among the points in a Latin hypercube design (LHD).

Sobol’, I. p. 2

Where the paper says it

, 2008; Da Veiga et al. , 2021) of which Sobol’ design (Sobol’, 2001) is a popular choice.

p. 2
Where the paper says it

, 2008; Da Veiga et al. , 2021) of which Sobol’ design (Sobol’, 2001) is a popular choice.

M., On sensitivity estimation for nonlinear mathematical models. p. 2
Where the paper says it

(2012) proposed to select a collection of constellations from an existing space-filling design. A first-order polynomial is then fitted within each of these local designs (simplex or constellation) to estimate the local sensitivity, but they can introduce bias in the estimation of global sensitivity indices when dealing with highly complex nonlinear models.

p. 2
Where the paper says it

(2012) proposed to select a collection of constellations from an existing space-filling design. A first-order polynomial is then fitted within each of these local designs (simplex or constellation) to estimate the local sensitivity, but they can introduce bias in the estimation of global sensitivity indices when dealing with highly complex nonlinear models.

Xiao, Q., Joseph, V. R., and Ray, D. M., Maximum one-factor-at-a-time designs for screening in computer experiments. p. 1

Where the paper says it

However, this sequential strategy can become infeasible for expensive computer experiments where all runs must be launched in a single batch to exploit parallel computing. Therefore, a one-shot design that supports both factor screening and surrogate modeling can be beneficial.

p. 1
Where the paper says it

However, this sequential strategy can become infeasible for expensive computer experiments where all runs must be launched in a single batch to exploit parallel computing. Therefore, a one-shot design that supports both factor screening and surrogate modeling can be beneficial.

Huang, C., Joseph, V. R., rkriging: Kriging Modeling. p. 0

Where the paper says it

Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology, Atlanta, GA 30332 October 7, 2026 Abstract Space-filling designs are commonly used in deterministic computer experiments. However, they cannot do factor screening without any modeling assumptions, which makes them inefficient when only a small subset of input factors is important.

Iooss, B., Veiga, S. p. 2

Where the paper says it

There are many Monte Carlo methods for estimating the global sensitivity indices (Saltelli et al. , 2008; Da Veiga et al.

D., Janon, A., and Pujol, G. p. 2
Where the paper says it

Although this strategy encourages OFATs to be placed as far apart as possible, it tends to push points toward the boundary of the input space, leaving large regions of the interior unexplored. Instead of fixating on OFAT designs, Pujol (2009) proposed using a collection of simplex designs to improve the projection properties.

, sensitivity: Global Sensitivity Analysis of Model Outputs and Importance Measures. p. 1
Where the paper says it

By identifying the important factors at an early stage, we can focus modeling efforts on the smaller active subspace, which can help us develop a more efficient and accurate surrogate model. This task is known as “factor screening” in sensitivity analysis (Saltelli et al.

Loeppky, J. L., Sacks, J., and Welch, W. J., Choosing the sample size of a computer experiment: A practical guide. p. 7

Where the paper says it

(4) Here, Aj and A (i) j denote the j-th row of A and A(i) , respectively. The sample variance var( c y) is given by Pn j=1(yj − y¯) 2/(n − 1), where ¯y = Pn j=1 yj/n is the sample mean and n = l × (p + 1) is the run size.

Marrel, A., Iooss, B. p. 6

Where the paper says it

Moreover, this strategy tends to push points toward the boundary, leaving large unexplored regions in the interior. Implementations of the Morris design and the optimal Morris design are available in the R package sensitivity (Iooss et al.

, and Chabridon, V., The ICSCREAM Methodology: Identification of Penalizing Configurations in Computer Experiments Using Screening and Metamodel—Applications in Thermal Hydraulics. p. 0
Where the paper says it

Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology, Atlanta, GA 30332 October 7, 2026 Abstract Space-filling designs are commonly used in deterministic computer experiments. However, they cannot do factor screening without any modeling assumptions, which makes them inefficient when only a small subset of input factors is important.

Cheng, S.-W., Factor screening and response surface exploration. p. 1

Where the paper says it

Space-filling designs are widely used as experimental designs in deterministic computer experiments (Joseph, 2026, Ch. These designs aim at placing points everywhere in the input space with as few gaps as possible, leading to robust performance for learning an unknown response surface.

Saltelli, A. p. 1

Where the paper says it

By identifying the important factors at an early stage, we can focus modeling efforts on the smaller active subspace, which can help us develop a more efficient and accurate surrogate model. This task is known as “factor screening” in sensitivity analysis (Saltelli et al.

, Ratto, M., Andres, T., Campolongo, F. p. 2
Where the paper says it

However, the random placement of OFATs can lead to poor coverage of the input space, as shown in Figure 1(a). To improve the coverage, Campolongo et al.

, Cariboni, J., Gatelli, D., Saisana, M., and Tarantola, S., Global sensitivity analysis: the primer. p. 1
Where the paper says it

By identifying the important factors at an early stage, we can focus modeling efforts on the smaller active subspace, which can help us develop a more efficient and accurate surrogate model. This task is known as “factor screening” in sensitivity analysis (Saltelli et al.

Sobol’, I. M., On sensitivity estimation for nonlinear mathematical models.

Song, D., Joseph, V. R., Efficient active learning strategies for computer experiments.

Sung, C.-L., Song, Y., and Hung, Y., Advancing inverse scattering with surrogate modeling and bayesian inference for functional inputs. p. 0

Where the paper says it

Keywords: Experimental design; Gaussian process; Screening designs; Space-filling designs; Surrogate modeling. 1 Introduction Computer models are essential tools for simulating complex systems, such as for optimizing materials crystal structures using density functional theory (Krishna et al.

Wang, S., Xie, W., and Joseph, V. p. 0

Where the paper says it

Space-Filling One-Factor-At-A-Time Designs Wei-Yang Yu and V. Roshan Joseph∗ H.

R., SFDesign: An R package for space-filling designs. p. 6
Where the paper says it

Implementations of the Morris design and the optimal Morris design are available in the R package sensitivity (Iooss et al. 2 MOFAT Designs Assuming that the underlying function between y and x is square integrable, Sobol’ (1990, 1993) shows that the variance of y can be decomposed into the variance resulting from the main effects, two-factor interactions, etc.

Xiao, Q., Joseph, V. R., and Ray, D. M., Maximum one-factor-at-a-time designs for screening in computer experiments.

Zhang, N., Apley, D. p. 8

Where the paper says it

equally spaced levels given by {0. To maximize the capacity of identifying important factors, A and B are chosen to maximize the expected total Sobol’ indices under the assumption that the responses are generated from a Brownian process (Zhang and Apley, 2014).

W., Fractional Brownian fields for response surface metamodeling. p. 1
Where the paper says it

Space-filling designs are widely used as experimental designs in deterministic computer experiments (Joseph, 2026, Ch. These designs aim at placing points everywhere in the input space with as few gaps as possible, leading to robust performance for learning an unknown response surface.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Develop a single-shot design that supports both factor screening and surrogate modeling, which can make sequential experiments more efficient when used as initial designs (Song and Joseph, 2026). This allows for simultaneous factor screening and surrogate model development without needing to fit a surrogate model first or imposing additional assumptions.

  2. Implement Space-filling One-Factor-at-a-Time (SOFT) designs that achieve optimal space-filling properties while retaining their screening capability. This design provides not only excellent coverage of the input space but also good screening ability due to its OFAT structure (Page 4).

  3. Develop Augmented SOFT (A-SOFT) designs that allocate part of the budget to space-filling points unrestricted by the OFAT structure. This improves flexibility for exploring the input space at the cost of fewer OFAT comparisons, providing a tradeoff between factor screening and space-fillingness (Page 5).

  4. Create a system that uses A-SOFT designs to achieve superior predictive performance over both space-filling designs and MOFAT design, especially in high-dimensional settings for complex functions like the Ackley function (Figure 11, Page 26).

  5. Enable the AI system to select active factors based on their total Sobol' indices estimated from data using either the direct OFAT structure (MOFAT and SOFT) or by fitting a Gaussian process model with the multiplicative inverse multiquadric (MIM) kernel (Page 4.3).

  6. Allow the AI to fit a Gaussian process model using only the factors identified as active after screening, simplifying the fitting task when an OFAT design is used (Song and Joseph, 2026).

Abstract

Space-filling designs are commonly used in deterministic computer experiments. However, they are ineffective for factor screening, which makes them inefficient when only a small subset of input factors is influential to the output. Recently developed screening designs, such as MOFAT designs, are effective at identifying important factors but lack space-filling properties, limiting their usefulness for surrogate modeling. In this article, we propose a new class of screening designs that improves the space-fillingness while retaining their screening capability. Through several numerical examples, we demonstrate that the proposed designs offer clear advantages over existing designs.

Related papers