Wednesday, November 18, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News

Neural Network Surrogate Accelerates Evolutionary Building Glazing Optimization Over 400-Fold

November 18, 2026
in Science News
Gavin Prescott
By Gavin Prescott Scienmag Editorial Profile - Ecology and Ecosystem Dynamics
Reading Time: 6 mins read
0
Neural Network Surrogate Accelerates Evolutionary Building Glazing Optimization Over 400-Fold

Neural Network Surrogate Accelerates Evolutionary Building Glazing Optimization Over 400-Fold

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

Integrating an artificial neural network as a computational surrogate within an evolutionary optimization framework accelerates the design of office building facade glazing by more than 400 times while identifying architectural configurations that achieve superior energy efficiency and lower construction costs compared to conventional simulation-driven searches. In a study published in Industrial Artificial Intelligence, researchers Alexander E. I. Brownlee and Ernest R. O. Vanmosuinck demonstrated that replacing time-intensive thermodynamic simulations with a trained multilayer perceptron enables multi-objective genetic algorithms to evaluate tens of thousands of candidate designs in minutes rather than weeks. The investigation focused on resolving the competing physical demands of daylight harvesting, thermal insulation, and capital expenditure across building envelopes, specifically targeting the complex combinatorial challenges inherent in binary-encoded facade layouts. By systematically training and deploying a surrogate model within the Non-Dominated Sorting Genetic Algorithm II, the research team established that surrogate-assisted evolutionary workflows can bypass computational bottlenecks that historically restricted whole-building numerical simulations in architectural design.

The physical envelope of a commercial building plays a decisive role in governing its operational energy profile, dictating how thermal gains and daylight interact with interior climate systems. Glazing layout directly influences heating, cooling, and artificial lighting demands through mechanisms that frequently conflict across changing seasonal conditions. Expansive window areas permit natural illumination to penetrate deep into interior zones, curtailing electricity consumption associated with artificial lighting. However, unshaded glazed apertures permit substantial solar radiation during warmer months, elevating cooling loads and air conditioning requirements. Conversely, during colder periods, extensive glass assemblies accelerate conductive heat losses across the exterior envelope, increasing heating demands. To model these physical tensions, the investigators examined a commercial atrium set within the climatic conditions of Chicago, Illinois, an environment characterized by extreme temperature swings between cold winters and hot summers. Balancing these environmental factors against construction costs requires identifying Pareto-optimal designs where operational energy use and capital cost cannot be reduced without compromising the other.

Translating building envelope design into an automated optimization framework requires reconciling high-dimensional architectural geometries with rigorous numerical simulation tools. The study modeled a southern atrium facade measuring 15 meters wide by 8.2 meters high, subdivided into an orthogonal grid of 120 discrete one-square-meter cells arranged fifteen wide by eight high. Each individual cell could be specified as either opaque concrete wall or double-glazed window, with an accompanying option to attach a physical shading overhang above any glazed cell to mitigate solar radiation. This design yielded a binary representation comprising 240 decision variables, where 120 bits governed cell glazing status and 120 bits controlled overhang deployment. Evaluating the primary performance objective—annual building energy consumption across heating, cooling, and lighting—required executing full-year thermodynamic simulations in EnergyPlus.

The computational expense of repeated EnergyPlus executions presents an immense barrier for evolutionary algorithms, which require thousands of evaluations to locate high-performing design regions. Conducting 50,000 evaluations using direct EnergyPlus simulations required roughly one full week of continuous processor runtime for a single optimization run, accumulating to 30 weeks across the thirty statistical replications needed for experimental confidence. While surrogate-assisted evolutionary algorithms have frequently been deployed across engineering disciplines to circumvent prohibitive simulation expenses, earlier applications within sustainable building design predominantly addressed continuous or mixed-integer variables, such as global window-to-wall ratios or insulation thicknesses. The researchers noted that binary-encoded facade problems, where discrete structural cells are toggled across a large grid, remained largely unaddressed by surrogate methodologies. Furthermore, previous architectural optimization studies typically selected a single machine learning model arbitrarily without systematically comparing alternative architectures or addressing the non-uniform distribution of design fitness across discrete binary spaces.

To overcome the limitations of prior implementations, the researchers developed a structured data-generation strategy tailored to binary facade encodings. In binary problems of this nature, the total count of active bits—corresponding to the total surface area of glazed cells—exerts a dominant influence over operational energy requirements and construction expenditures. Standard uniform random sampling across a 240-bit space overwhelmingly generates candidate configurations clustered tightly around a 50 percent glazing ratio, leaving extreme configurations near completely opaque or fully glazed envelopes unrepresented. To ensure comprehensive coverage across the search landscape, the authors formulated a specialized sampling algorithm that generated 500 training solutions evenly distributed across the spectrum of glazing fractions, spanning from zero to one hundred percent glazed facade areas. Each of these 500 representative configurations was subsequently evaluated through EnergyPlus simulations to establish an authoritative benchmark dataset for machine learning model training and hyperparameter optimization.

The authors subsequently conducted a systematic comparative evaluation of multiple machine learning architectures using fivefold cross-validation on the structured 500-solution training dataset. Model configurations were tuned through random searches spanning 500 hyperparameter combinations within the WEKA machine learning suite, assessing candidate algorithms by their coefficient of determination, or r-squared value. The multilayer perceptron emerged as the most accurate surrogate, demonstrating an average r-squared value of 0.992 across unseen cross-validation folds despite training on a dataset representing an infinitesimally small fraction of the vast theoretical search space. Hyperparameter tuning demonstrated that the standard multilayer perceptron configuration—featuring a single hidden layer of 240 neurons matching the input feature count, a learning rate of 0.3, a momentum of 0.2, and 500 training epochs—provided robust generalization without overfitting. Because the target evolutionary algorithm uses tournament selection, the surrogate was primarily required to preserve relative solution rankings rather than absolute numerical precision, making high correlation across predictions particularly advantageous.

The trained multilayer perceptron was directly integrated into the Non-Dominated Sorting Genetic Algorithm II, replacing the EnergyPlus simulation engine during the internal evolutionary search loop. The optimization sought to simultaneously minimize annual energy consumption and capital construction costs, the latter calculated via a deterministic linear cost function assigning 350 pounds per glazed cell, 112 pounds per opaque cell, and 128 pounds per shading overhang. To assess whether tracking dynamic evolutionary trajectories required adaptive updating, the researchers formulated two surrogate configurations: a static protocol where the initial neural network remained unchanged throughout the run, and an iterative retraining protocol where the model was retrained from scratch at quartile intervals using newly evaluated population members.

To quantify performance, thirty independent experimental runs were conducted across three algorithm configurations limited to 50,000 evaluations: standard NSGA-II relying exclusively on EnergyPlus, surrogate-assisted NSGA-II without retraining, and surrogate-assisted NSGA-II with periodic quartile retraining. Progress was evaluated using empirical attainment surfaces and normalized hypervolume indicators bounded against a reference coordinate of unity. In addition to dramatic computational savings, both surrogate variants discovered superior Pareto fronts compared to the unassisted algorithm. The median run of the simulation-only algorithm located trade-off solutions reaching 43,746 kilowatt-hours of annual energy at 18,500 pounds capital expenditure. In contrast, the static surrogate discovered solutions reaching 41,611 kilowatt-hours at 17,750 pounds, while the retrained surrogate achieved 41,458 kilowatt-hours at 17,750 pounds, demonstrating that rapid surrogate evaluations enabled deeper exploration of promising trade-off boundaries.

Capitalizing on the extreme speedup enabled by surrogate evaluation, the research team conducted extended optimization experiments extending to 100,000 search iterations, a computational depth that would require an estimated 104 days under standard EnergyPlus simulations. Within this expanded horizon, the static surrogate finished median runs in 477 seconds, whereas the periodically retrained surrogate concluded in 702 seconds due to the modest overhead of intermediate simulation calls. The extended search yielded marked enhancements in solution quality across the Pareto frontier, with periodic model retraining providing a distinct advantage in the energy domain. The non-updated surrogate achieved median solutions reaching 41,072 kilowatt-hours at 15,750 pounds, whereas the retrained surrogate achieved 40,420 kilowatt-hours at 16,000 pounds. The authors observed that because capital cost was calculated deterministically without surrogate estimation, errors in the neural network affected only energy predictions. Periodic retraining corrected these accumulated surrogate biases over successive generations, preventing the search from diverging into sub-optimal regions.

Post-experimental analysis provided crucial insights into surrogate model behavior and design morphology across the discovered Pareto fronts. Inspection of top-performing facades revealed that glazed cells dispersed evenly across the wall rather than forming contiguous clusters, indicating that overall glazing volume exerted far greater influence on thermal performance than localized geometric grouping. Evaluating the surrogate models against full EnergyPlus recalculations of the final 100,000-iteration population revealed Spearman rank correlation coefficients of 0.368 for the static model and 0.407 for the retrained model. Although absolute prediction error expanded noticeably for heavily glazed facades with numerous shading overhangs, the surrogate retained sufficient rank-ordering fidelity to guide tournament selection reliably toward high-performing designs. The investigators concluded that structured training data and periodic model retraining provide an efficient, robust framework for deploying machine learning surrogates in binary architectural optimization, establishing a foundation for future inquiries into adaptive retraining triggers and alternative discrete encodings that eliminate redundant shading variables.

Subject of Research: Surrogate-assisted evolutionary multi-objective optimization of building facade glazing

Article Title: Surrogate-assisted evolutionary multi-objective optimisation of office building glazing

Article References: Brownlee, A. E. I., & Vanmosuinck, E. R. O. (2025). Surrogate-assisted evolutionary multi-objective optimisation of office building glazing. Industrial Artificial Intelligence, 3(1), Article 4. https://doi.org/10.1007/s44244-025-00025-1

Image Credits: AI Generated

DOI: 10.1007/s44244-025-00025-1

Keywords: surrogate modeling, evolutionary algorithms, NSGA-II, building performance simulation, EnergyPlus, facade glazing, multi-objective optimization, multilayer perceptron, architectural engineering, building energy efficiency, Pareto optimality, binary encoding, shading overhangs, hypervolume indicator, construction cost minimization, computational design, machine learning in engineering

Cite Scienmag News

Gavin Prescott. (November 18, 2026). Neural Network Surrogate Accelerates Evolutionary Building Glazing Optimization Over 400-Fold. Scienmag. https://scienmag.com/neural-network-surrogate-accelerates-evolutionary-building-glazing-optimization-over-400-fold/

Gavin Prescott. "Neural Network Surrogate Accelerates Evolutionary Building Glazing Optimization Over 400-Fold." Scienmag, 18 November 2026, https://scienmag.com/neural-network-surrogate-accelerates-evolutionary-building-glazing-optimization-over-400-fold/. Accessed 18 November 2026.

Gavin Prescott. "Neural Network Surrogate Accelerates Evolutionary Building Glazing Optimization Over 400-Fold." Scienmag. November 18, 2026. https://scienmag.com/neural-network-surrogate-accelerates-evolutionary-building-glazing-optimization-over-400-fold/

Tags: architectural engineeringbinary encodingbuilding energy efficiencybuilding performance simulationcomputational designconstruction cost minimizationEnergyPlusevolutionary algorithmsfacade glazinghypervolume indicatormachine learning in engineeringmulti-objective optimizationmultilayer perceptronNSGA-IIPareto optimalityshading overhangssurrogate modeling
Share26Tweet16
Previous Post

Polygenic Risk Scores Show Limited Cross-Population Portability Across Two Biobanks

Next Post

High-Achieving University Students Adopt Generative AI Guardedly to Protect Learning Autonomy

Related Posts

Higher Plant Density Boosts Blue Panicum Forage Yield Across Saline Lands
Agriculture

Higher Plant Density Boosts Blue Panicum Forage Yield Across Saline Lands

November 18, 2026
Cumulative Vascular Compromise Drives Rare Maxillary Necrosis After Le Fort Osteotomy
Medicine

Cumulative Vascular Compromise Drives Rare Maxillary Necrosis After Le Fort Osteotomy

November 18, 2026
Community Waste Framework Balances Tourism Pressures on Indonesia’s Sabang Island
Science News

Community Waste Framework Balances Tourism Pressures on Indonesia’s Sabang Island

November 18, 2026
Price Fairness and Service Quality Drive Aesthetic Clinic Repurchase Through Brand Trust
Science News

Price Fairness and Service Quality Drive Aesthetic Clinic Repurchase Through Brand Trust

November 18, 2026
Human Slow-Release Buprenorphine Provides Sustained Pain Relief in Laboratory Rats
Biology

Human Slow-Release Buprenorphine Provides Sustained Pain Relief in Laboratory Rats

November 18, 2026
Novel SERPING1 Gene Mutation Identified in Chinese Family With Hereditary Angioedema
Science News

Novel SERPING1 Gene Mutation Identified in Chinese Family With Hereditary Angioedema

November 18, 2026
Next Post
High-Achieving University Students Adopt Generative AI Guardedly to Protect Learning Autonomy

High-Achieving University Students Adopt Generative AI Guardedly to Protect Learning Autonomy

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Higher Plant Density Boosts Blue Panicum Forage Yield Across Saline Lands
  • Cumulative Vascular Compromise Drives Rare Maxillary Necrosis After Le Fort Osteotomy
  • Community Waste Framework Balances Tourism Pressures on Indonesia’s Sabang Island
  • Price Fairness and Service Quality Drive Aesthetic Clinic Repurchase Through Brand Trust

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading