Fan Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Fan Zhang has a total of 113 co-authors. Here are the top co-authors by number of shared publications:
Samee U. Khan
Mississippi State University
7 shared publicationsYifang Luo
6 shared publicationsYihe Wang
State Key Laboratory of Ocean Sensing & Ocean College, Zhejiang University, Zhoushan, 316021, China
5 shared publications- Bohua Li
Guangxi University
4 shared publications Hang Diao
3 shared publications- Joel Meyers
Southern Methodist University
3 shared publications
Latest Publications
- Open Access
- 26 October 2026
- Published by IOP Publishing in Machine Learning: Science and Technology
A provenance-aware zirconia conductivity database and Arrhenius-constrained screening of doped electrolyte candidates
Zirconia solid electrolytes are key oxygen-ion conductors, but reported conductivity values depend strongly on dopant chemistry, processing history and measurement temperature. Here we curate a provenance-aware zirconia conductivity database of 1,351 records and use it to compare an Arrhenius-constrained physics-informed machine-learning model (PIML) with an unconstrained deep neural network (DNN). In a reserved grouped-material confirmation split, the DNN achieves the lower pooled root-mean-square error (RMSE, in log 10 σ units) for withheld composition-processing groups. The Arrhenius-constrained model remains on a comparable error scale and provides bounded effective activation-energy estimates. Its clearest benefit appears in matched-material temperature extrapolation: when trained at ≤ 800 • C and tested at ≥ 900 • C (n = 47), PIML reduces RMSE from 0.4187 to 0.3609 and raises R 2 from 0.8562 to 0.8932. Bootstrap analysis supports this record-level gain, with material-fingerprint resampling reported as a conservative uncertainty check. Genetic-algorithm (GA) screening under fixed processing and measurement conditions identifies a Sc-rich/Mg-containing zirconia region. Exploratory CHGNet molecular dynamics provides the main computational ranking check, with single-configuration oxide-reference density functional theory (DFT) retained as an auxiliary thermodynamic screen. This work provides a curated conductivity resource and a conservative screening route for doped zirconia electrolytes.
- 26 October 2026
- Published by IOP Publishing in Machine Learning: Science and Technology
Open AccessA provenance-aware zirconia conductivity database and Arrhenius-constrained screening of doped electrolyte candidates
Zirconia solid electrolytes are key oxygen-ion conductors, but reported conductivity values depend strongly on dopant chemistry, processing history and measurement temperature. Here we curate a provenance-aware zirconia conductivity database of 1,351 records and use it to compare an Arrhenius-constrained physics-informed machine-learning model (PIML) with an unconstrained deep neural network (DNN). In a reserved grouped-material confirmation split, the DNN achieves the lower pooled root-mean-square error (RMSE, in log 10 σ units) for withheld composition-processing groups. The Arrhenius-constrained model remains on a comparable error scale and provides bounded effective activation-energy estimates. Its clearest benefit appears in matched-material temperature extrapolation: when trained at ≤ 800 • C and tested at ≥ 900 • C (n = 47), PIML reduces RMSE from 0.4187 to 0.3609 and raises R 2 from 0.8562 to 0.8932. Bootstrap analysis supports this record-level gain, with material-fingerprint resampling reported as a conservative uncertainty check. Genetic-algorithm (GA) screening under fixed processing and measurement conditions identifies a Sc-rich/Mg-containing zirconia region. Exploratory CHGNet molecular dynamics provides the main computational ranking check, with single-configuration oxide-reference density functional theory (DFT) retained as an auxiliary thermodynamic screen. This work provides a curated conductivity resource and a conservative screening route for doped zirconia electrolytes.
- Open Access
- 31 August 2026
- Published by MDPI AG in Mathematics
- Vol. 14 (17) , 3130
Equal Budgets Change the Verdict: Finite-Sample Bias and a Matched-Budget Re-Examination of Diversity-Enhanced flowMC Ensembles
Normalizing-flow Markov chain Monte Carlo (MCMC), such as flowMC, augments local moves with a learned global flow proposal; a natural reliability idea is to pool samples from several such samplers. On 6 targets in 20 and 50 dimensions, a 5-member diverse flowMC ensemble appeared to reduce the average marginal Jensen–Shannon (JS) distance by about 18% relative to a single flowMC run. Matching the returned sample counts—our “equal budget”: wall-clock costs differ and are reported separately—reverses this verdict: the ensemble draws five times as many samples, and even a perfect sampler’s histogram JS estimate has a closed root-mean-square finite-sample floor , with c a fitted coefficient set by the binning, which explains the apparent gain almost entirely. At a matched 10,000-sample budget, a single flowMC run matches or beats this ensemble on all eleven configurations and runs about six times faster. Applied to exact draws, the uncorrected reweight-and-resample aggregation reproduces about 93% of the ensemble’s elevation above the floor, separating a genuine law shift (coverage and tempering reweighting) from an estimator-level loss (resampling). On two real Bayesian posteriors scored against long NUTS references, uniform pooling improved on the sample-count-matched single run under marginal JS, energy distance, and in every recorded run (a descriptive comparison at three and two repetitions), while the aggregation’s deficit is confined to the biased marginal metric. We recommend matched sample counts and floor reporting for histogram divergences, unbiased joint metrics alongside, and uniform pooling instead of uncorrected reweighting and resampling.- 31 August 2026
- Published by MDPI AG in Mathematics
- Vol. 14 (17) , 3130
Open AccessEqual Budgets Change the Verdict: Finite-Sample Bias and a Matched-Budget Re-Examination of Diversity-Enhanced flowMC Ensembles
Normalizing-flow Markov chain Monte Carlo (MCMC), such as flowMC, augments local moves with a learned global flow proposal; a natural reliability idea is to pool samples from several such samplers. On 6 targets in 20 and 50 dimensions, a 5-member diverse flowMC ensemble appeared to reduce the average marginal Jensen–Shannon (JS) distance by about 18% relative to a single flowMC run. Matching the returned sample counts—our “equal budget”: wall-clock costs differ and are reported separately—reverses this verdict: the ensemble draws five times as many samples, and even a perfect sampler’s histogram JS estimate has a closed root-mean-square finite-sample floor , with c a fitted coefficient set by the binning, which explains the apparent gain almost entirely. At a matched 10,000-sample budget, a single flowMC run matches or beats this ensemble on all eleven configurations and runs about six times faster. Applied to exact draws, the uncorrected reweight-and-resample aggregation reproduces about 93% of the ensemble’s elevation above the floor, separating a genuine law shift (coverage and tempering reweighting) from an estimator-level loss (resampling). On two real Bayesian posteriors scored against long NUTS references, uniform pooling improved on the sample-count-matched single run under marginal JS, energy distance, and in every recorded run (a descriptive comparison at three and two repetitions), while the aggregation’s deficit is confined to the biased marginal metric. We recommend matched sample counts and floor reporting for histogram divergences, unbiased joint metrics alongside, and uniform pooling instead of uncorrected reweighting and resampling. - 18 August 2026
- Published by American Physical Society (APS) in Physical Review D
- Vol. 114 (4) , 043535
Accurate neural network emulator for primordial light element abundances
Big bang nucleosynthesis (BBN) predictions of primordial light-element abundances offer a powerful probe of early-Universe physics. However, high-accuracy numerical BBN calculations have become a major computational bottleneck for large-scale cosmological inferences due to the complex nuclear network. Here we present bbnet, a fast and accurate deep learning emulator for primordial abundances. The training data are generated by full numerical calculations using two public BBN codes, parthenope and alterbbn, modified to accommodate extended cosmologies that include dark radiation and a stiff equation of state. The network employs a residual multihead architecture to capture convoluted physical relationships. bbnet produces primordial helium-4 and deuterium abundances with negligible errors in milliseconds per sample, achieving a speed-up of up to ${10}^{4}$ times relative to first-principles solvers while remaining unbiased over wide parameter ranges. Therefore, our emulator can supersede traditional simplified numerical prescriptions that compromise accuracy for speed. Based on extensive assessments of its performance, we conclude that bbnet is an optimal solution to the theoretical prediction of primordial element abundances. It will serve as a reliable tool for precision cosmology and new physics searches.- 18 August 2026
- Published by American Physical Society (APS) in Physical Review D
- Vol. 114 (4) , 043535
Accurate neural network emulator for primordial light element abundances
Big bang nucleosynthesis (BBN) predictions of primordial light-element abundances offer a powerful probe of early-Universe physics. However, high-accuracy numerical BBN calculations have become a major computational bottleneck for large-scale cosmological inferences due to the complex nuclear network. Here we present bbnet, a fast and accurate deep learning emulator for primordial abundances. The training data are generated by full numerical calculations using two public BBN codes, parthenope and alterbbn, modified to accommodate extended cosmologies that include dark radiation and a stiff equation of state. The network employs a residual multihead architecture to capture convoluted physical relationships. bbnet produces primordial helium-4 and deuterium abundances with negligible errors in milliseconds per sample, achieving a speed-up of up to ${10}^{4}$ times relative to first-principles solvers while remaining unbiased over wide parameter ranges. Therefore, our emulator can supersede traditional simplified numerical prescriptions that compromise accuracy for speed. Based on extensive assessments of its performance, we conclude that bbnet is an optimal solution to the theoretical prediction of primordial element abundances. It will serve as a reliable tool for precision cosmology and new physics searches.- Open Access
- 11 August 2026
- Published by Copernicus GmbH
- p. 1-24
Comment on egusphere-2026-2115
As one of the fundamental mechanical properties of snow, the elastic modulus is critical to the design and construction of Antarctic snow runways and roads. While previous studies measured and investigated snow’s elastic modulus through various experimental methods, the effects of vibratory treatment, a construction measure proven to enhance snow hardness, on the elastic modulus and the underlying mesoscale mechanism remain unexamined. This study investigates the vibration-induced effect on elastic modulus of compacted Antarctic snow near Zhongshan Station and the corresponding mesoscale mechanism. P-wave propagation experiments were conducted to measure the elastic modulus of vibrated and non-vibrated compacted Antarctic snow, and X-ray tomography imaging was employed to obtain the microstructures of vibrated and non-vibrated snow samples. Results show that for isothermal sintering of 48 hours at -10 °C, vibratory treatment increases the elastic modulus by 83.13 % while maintaining the snow density of 0.6 g/cm³ unchanged. At the mesoscale, vibratory treatment effectively homogenizes the pore space distribution within the ice matrix. Quantitative analysis revealed the following microstructural modifications: a 7.14 % decrease in the mean structure thickness accompanied by a 12.41 % reduction in the standard deviation, a 13.68 % decrease in the mean pore thickness with a more pronounced 30.43 % decline in the standard deviation, an 18.82 % elevation in the minimum cut density, and a 2.09 % enhancement in the directional connectivity. The findings provide theoretical support for rapid construction techniques of Antarctic snow runways and roads.
- 11 August 2026
- Published by Copernicus GmbH
- p. 1-24
Open AccessComment on egusphere-2026-2115
As one of the fundamental mechanical properties of snow, the elastic modulus is critical to the design and construction of Antarctic snow runways and roads. While previous studies measured and investigated snow’s elastic modulus through various experimental methods, the effects of vibratory treatment, a construction measure proven to enhance snow hardness, on the elastic modulus and the underlying mesoscale mechanism remain unexamined. This study investigates the vibration-induced effect on elastic modulus of compacted Antarctic snow near Zhongshan Station and the corresponding mesoscale mechanism. P-wave propagation experiments were conducted to measure the elastic modulus of vibrated and non-vibrated compacted Antarctic snow, and X-ray tomography imaging was employed to obtain the microstructures of vibrated and non-vibrated snow samples. Results show that for isothermal sintering of 48 hours at -10 °C, vibratory treatment increases the elastic modulus by 83.13 % while maintaining the snow density of 0.6 g/cm³ unchanged. At the mesoscale, vibratory treatment effectively homogenizes the pore space distribution within the ice matrix. Quantitative analysis revealed the following microstructural modifications: a 7.14 % decrease in the mean structure thickness accompanied by a 12.41 % reduction in the standard deviation, a 13.68 % decrease in the mean pore thickness with a more pronounced 30.43 % decline in the standard deviation, an 18.82 % elevation in the minimum cut density, and a 2.09 % enhancement in the directional connectivity. The findings provide theoretical support for rapid construction techniques of Antarctic snow runways and roads.
- Open Access
- 25 July 2026
- Published by Elsevier in SSRN Electronic Journal
Navigation-Risk-Aware Ordinal Segmentation for SAR Sea-Ice Stage-of-Development Mapping
Sea-ice stage-of-development (SOD) mapping from synthetic aperture radar (SAR) is a key input for polar navigation safety. Existing deep-learning methods treat SOD as ordinary multi-class segmentation and optimize accuracy alone. However, SOD classes are intrinsically ordinal, from open water to multi-year ice, and their errors are asymmetric for navigation: underestimating ice severity is far more dangerous than overestimating it. We propose a navigation-risk-aware ordinal segmentation framework that makes this asymmetry an explicit learning target. An asymmetric navigation cost matrix encodes ordinal distance and underestimation risk; complementary safety metrics expose the operational risk that accuracy cannot capture; and the cost matrix is injected into both the loss and the inference decision, comparing matrix-expected-risk, asymmetric-soft-label, ordinal Brier and CORAL losses against Bayes-minimum-risk and CVaR decoding on a shared SAR backbone. On a Sentinel-1 dataset, a lightweight cost-aware auxiliary loss, CE+0.2 MERL, Pareto-dominates both a standard cross-entropy baseline and a strong cost-aware Bayes-risk decoder, improving F1 by 1.19 points and mIoU by 1.51 points while reducing expected navigation cost by ~28.2% and the underestimation rate by ~40.4%. Encoding safety asymmetry into training rather than only into the decision rule yields models that are both more accurate and safer for operational ice charting.
- 25 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessNavigation-Risk-Aware Ordinal Segmentation for SAR Sea-Ice Stage-of-Development Mapping
Sea-ice stage-of-development (SOD) mapping from synthetic aperture radar (SAR) is a key input for polar navigation safety. Existing deep-learning methods treat SOD as ordinary multi-class segmentation and optimize accuracy alone. However, SOD classes are intrinsically ordinal, from open water to multi-year ice, and their errors are asymmetric for navigation: underestimating ice severity is far more dangerous than overestimating it. We propose a navigation-risk-aware ordinal segmentation framework that makes this asymmetry an explicit learning target. An asymmetric navigation cost matrix encodes ordinal distance and underestimation risk; complementary safety metrics expose the operational risk that accuracy cannot capture; and the cost matrix is injected into both the loss and the inference decision, comparing matrix-expected-risk, asymmetric-soft-label, ordinal Brier and CORAL losses against Bayes-minimum-risk and CVaR decoding on a shared SAR backbone. On a Sentinel-1 dataset, a lightweight cost-aware auxiliary loss, CE+0.2 MERL, Pareto-dominates both a standard cross-entropy baseline and a strong cost-aware Bayes-risk decoder, improving F1 by 1.19 points and mIoU by 1.51 points while reducing expected navigation cost by ~28.2% and the underestimation rate by ~40.4%. Encoding safety asymmetry into training rather than only into the decision rule yields models that are both more accurate and safer for operational ice charting.
- 17 July 2026
- Published by Springer Nature
- p. 81-95
Automated Injury Severity Assessment Using Knowledge-Grounded Large Language Models
- 17 July 2026
- Published by Springer Nature
- p. 81-95
Automated Injury Severity Assessment Using Knowledge-Grounded Large Language Models
- Open Access
- 19 May 2026
- Published by IOP Publishing in Machine Learning: Science and Technology
- Vol. 7 (3) , 035023
MARVEL: a multi-agent research validator and enabler using large language models
We present MARVEL ( https://ligogpt.mit.edu/marvel ), a locally deployable, open-source framework for domain-aware question answering and assisted scientific research. This system has been developed to work with technical data from scientific groups and to provide citation-backed responses while operating in authenticated computing environments. For complex queries, MARVEL uses DeepSearch, which integrates retrieval-augmented generation with Monte Carlo Tree Search. The DeepSearch mode breaks a query into related sub-questions and spends more compute on branches that appear useful, while keeping track of sources through a global evidence ledger during drafting. We apply the framework to gravitational-wave (GW) research, using material related to the laser Interferometer GW Observatory (LIGO). Answers are grounded in a curated semantic index of research literature, doctoral theses, LIGO documents, and long-running detector electronic logbooks, with targeted web searches when appropriate. Since commercial large language models cannot be directly benchmarked on private data, we evaluate MARVEL using two publicly available datasets chosen to resemble the semantic and technical characteristics of our target domain. On these datasets, results for literature-style queries are comparable between MARVEL and a GPT-4o mini baseline. We see improved results in queries related to detector operations where the effects of domain-specific retrieval and multi-step reasoning are more apparent. The code and evaluation datasets are released with this work.
- 19 May 2026
- Published by IOP Publishing in Machine Learning: Science and Technology
- Vol. 7 (3) , 035023
Open AccessMARVEL: a multi-agent research validator and enabler using large language models
We present MARVEL ( https://ligogpt.mit.edu/marvel ), a locally deployable, open-source framework for domain-aware question answering and assisted scientific research. This system has been developed to work with technical data from scientific groups and to provide citation-backed responses while operating in authenticated computing environments. For complex queries, MARVEL uses DeepSearch, which integrates retrieval-augmented generation with Monte Carlo Tree Search. The DeepSearch mode breaks a query into related sub-questions and spends more compute on branches that appear useful, while keeping track of sources through a global evidence ledger during drafting. We apply the framework to gravitational-wave (GW) research, using material related to the laser Interferometer GW Observatory (LIGO). Answers are grounded in a curated semantic index of research literature, doctoral theses, LIGO documents, and long-running detector electronic logbooks, with targeted web searches when appropriate. Since commercial large language models cannot be directly benchmarked on private data, we evaluate MARVEL using two publicly available datasets chosen to resemble the semantic and technical characteristics of our target domain. On these datasets, results for literature-style queries are comparable between MARVEL and a GPT-4o mini baseline. We see improved results in queries related to detector operations where the effects of domain-specific retrieval and multi-step reasoning are more apparent. The code and evaluation datasets are released with this work.
- Open Access
- 12 May 2026
- Published by Copernicus GmbH in EGUsphere
Effect of vibration on the elastic modulus of compacted Antarctic snow near Zhongshan Station
As one of the fundamental mechanical properties of snow, the elastic modulus is critical to the design and construction of Antarctic snow runways and roads. While previous studies measured and investigated snow’s elastic modulus through various experimental methods, the effects of vibratory treatment, a construction measure proven to enhance snow hardness, on the elastic modulus and the underlying mesoscale mechanism remain unexamined. This study investigates the vibration-induced effect on elastic modulus of compacted Antarctic snow near Zhongshan Station and the corresponding mesoscale mechanism. P-wave propagation experiments were conducted to measure the elastic modulus of vibrated and non-vibrated compacted Antarctic snow, and X-ray tomography imaging was employed to obtain the microstructures of vibrated and non-vibrated snow samples. Results show that for isothermal sintering of 48 hours at -10 °C, vibratory treatment increases the elastic modulus by 83.13 % while maintaining the snow density of 0.6 g/cm³ unchanged. At the mesoscale, vibratory treatment effectively homogenizes the pore space distribution within the ice matrix. Quantitative analysis revealed the following microstructural modifications: a 7.14 % decrease in the mean structure thickness accompanied by a 12.41 % reduction in the standard deviation, a 13.68 % decrease in the mean pore thickness with a more pronounced 30.43 % decline in the standard deviation, an 18.82 % elevation in the minimum cut density, and a 2.09 % enhancement in the directional connectivity. The findings provide theoretical support for rapid construction techniques of Antarctic snow runways and roads.
- 12 May 2026
- Published by Copernicus GmbH in EGUsphere
Open AccessEffect of vibration on the elastic modulus of compacted Antarctic snow near Zhongshan Station
As one of the fundamental mechanical properties of snow, the elastic modulus is critical to the design and construction of Antarctic snow runways and roads. While previous studies measured and investigated snow’s elastic modulus through various experimental methods, the effects of vibratory treatment, a construction measure proven to enhance snow hardness, on the elastic modulus and the underlying mesoscale mechanism remain unexamined. This study investigates the vibration-induced effect on elastic modulus of compacted Antarctic snow near Zhongshan Station and the corresponding mesoscale mechanism. P-wave propagation experiments were conducted to measure the elastic modulus of vibrated and non-vibrated compacted Antarctic snow, and X-ray tomography imaging was employed to obtain the microstructures of vibrated and non-vibrated snow samples. Results show that for isothermal sintering of 48 hours at -10 °C, vibratory treatment increases the elastic modulus by 83.13 % while maintaining the snow density of 0.6 g/cm³ unchanged. At the mesoscale, vibratory treatment effectively homogenizes the pore space distribution within the ice matrix. Quantitative analysis revealed the following microstructural modifications: a 7.14 % decrease in the mean structure thickness accompanied by a 12.41 % reduction in the standard deviation, a 13.68 % decrease in the mean pore thickness with a more pronounced 30.43 % decline in the standard deviation, an 18.82 % elevation in the minimum cut density, and a 2.09 % enhancement in the directional connectivity. The findings provide theoretical support for rapid construction techniques of Antarctic snow runways and roads.
- Open Access
- 4 March 2026
- Published by Springer Nature in Research Square
Efficient self-attention-based emulation of gravitational-wave spectrum with full parameter space exploration
Fast emulation of sequential data generated by numerical computations is vital for large-scale Bayesian analysis in cosmology and astrophysics. Despite the success of neural network models in such tasks, benchmarks typically fall short of the required sub-percent accuracy and lack generalizability to broader data sets of different features and physical origins. Here we introduce SageNet+, a physics-informed framework for emulating gravitational-wave spectrum, while our self-attention-based design is adaptable to generic sequential data. First-principles modeling of the stochastic gravitational-wave background from cosmic inflation takes tens of seconds per evaluation to solve differential integrations. Instead, our SageNet+ emulator can predict the present-day gravitational-wave spectrum in less than 10 ms for every set of the nine cosmological parameters with wide prior ranges, achieving a thousand-fold speed-up over numerical methods. Based on the Transformer architecture, SageNet+ combines learnable positional embeddings and adaptive sequence standardization to preserve spectral features. We further employ refined data preprocessing to maximize physical information and devise novel metrics to assess model performance. Trained on 180,000 spectra that explore the full parameter space, SageNet+ attains sub-percent accuracy for all test data. This enables real-time parameter inferences and joint analyses with external data. SageNet+ presents an optimal, systematic and generalizable framework for emulation tasks in physical sciences.
- 4 March 2026
- Published by Springer Nature in Research Square
Open AccessEfficient self-attention-based emulation of gravitational-wave spectrum with full parameter space exploration
Fast emulation of sequential data generated by numerical computations is vital for large-scale Bayesian analysis in cosmology and astrophysics. Despite the success of neural network models in such tasks, benchmarks typically fall short of the required sub-percent accuracy and lack generalizability to broader data sets of different features and physical origins. Here we introduce SageNet+, a physics-informed framework for emulating gravitational-wave spectrum, while our self-attention-based design is adaptable to generic sequential data. First-principles modeling of the stochastic gravitational-wave background from cosmic inflation takes tens of seconds per evaluation to solve differential integrations. Instead, our SageNet+ emulator can predict the present-day gravitational-wave spectrum in less than 10 ms for every set of the nine cosmological parameters with wide prior ranges, achieving a thousand-fold speed-up over numerical methods. Based on the Transformer architecture, SageNet+ combines learnable positional embeddings and adaptive sequence standardization to preserve spectral features. We further employ refined data preprocessing to maximize physical information and devise novel metrics to assess model performance. Trained on 180,000 spectra that explore the full parameter space, SageNet+ attains sub-percent accuracy for all test data. This enables real-time parameter inferences and joint analyses with external data. SageNet+ presents an optimal, systematic and generalizable framework for emulation tasks in physical sciences.
- Open Access
- 26 January 2026
- Published by MDPI AG in Algorithms
- Vol. 19 (2) , 98
Machine Learning Applications for Sustainable Housing Policy: Understanding Price Determinants to Inform Affordable Housing Strategies
Understanding how housing attributes are capitalized into prices is central to addressing urban affordability challenges. Using 2799 second-hand housing transactions from Wenzhou, China, this study examines residential price formation under pronounced spatial and structural heterogeneity. Multiple predictive models are evaluated within a unified 10-fold cross-validation framework. Results indicate that Random Forest delivers the strongest predictive performance, achieving a normalized mean squared error below 0.10 and explaining over 90% of out-of-sample price variation, substantially outperforming hedonic regression, regression trees, bagging, boosting, and support vector models. Permutation-based importance analysis identifies district location, building scale, and floor area as the dominant price determinants, while the influence of renovation quality, transportation access, and educational amenities varies across districts and dwelling types. These findings reveal strong nonlinearities and heterogeneous valuation mechanisms in rapidly urbanizing housing markets. Methodologically, the study demonstrates how interpretable machine learning complements traditional hedonic analysis, while providing policy-relevant insights into housing affordability dynamics in medium-sized Chinese cities.
- 26 January 2026
- Published by MDPI AG in Algorithms
- Vol. 19 (2) , 98
Open AccessMachine Learning Applications for Sustainable Housing Policy: Understanding Price Determinants to Inform Affordable Housing Strategies
Understanding how housing attributes are capitalized into prices is central to addressing urban affordability challenges. Using 2799 second-hand housing transactions from Wenzhou, China, this study examines residential price formation under pronounced spatial and structural heterogeneity. Multiple predictive models are evaluated within a unified 10-fold cross-validation framework. Results indicate that Random Forest delivers the strongest predictive performance, achieving a normalized mean squared error below 0.10 and explaining over 90% of out-of-sample price variation, substantially outperforming hedonic regression, regression trees, bagging, boosting, and support vector models. Permutation-based importance analysis identifies district location, building scale, and floor area as the dominant price determinants, while the influence of renovation quality, transportation access, and educational amenities varies across districts and dwelling types. These findings reveal strong nonlinearities and heterogeneous valuation mechanisms in rapidly urbanizing housing markets. Methodologically, the study demonstrates how interpretable machine learning complements traditional hedonic analysis, while providing policy-relevant insights into housing affordability dynamics in medium-sized Chinese cities.