Data-efficient flood depth prediction through domain-aware coreset selection and tabular foundation models
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Data-efficient flood depth prediction through domain-aware coreset selection and tabular foundation models".
Jane: Here's the gist of the paper. Data-efficient flood depth prediction through domain-aware coreset selection and tabular foundation models is proposed to provide accurate, fast, and transferable watershed-level flood depth predictions by conditioning a tabular foundation model at inference time using a domain-aware coreset <ref:2606.05265#pg2> This pipeline achieves a mean R2 of 0.663 across nine Houston-area watersheds with only 0.7% of the per-watershed training pool <ref:2606.05265#pg9> It demonstrates that a vanilla TFM conditioned on the domain-aware coreset recovers 98.5% of a supervised baseline on held-out watersheds without retraining <ref:2606.05265#pg2> Domain-Aware Coreset Construction The pipeline involves a two-stage construction process to compress the physics-based simulation archive into a compact, hydrologically representative subset <ref:2606.05265#pg4> Stage 1 stratifies storm events by return period and most-affected watershed, while Stage 2 selects hexagons with a target-aware spatial selector <ref:2606.05265#pg5> The stratification ensures that rare high-RP events and their host watersheds survive sampling noise <ref:2606.05265#pg2> The coreset is the Cartesian product of N events sampled in Stage 1 and Nh hexagons selected per watershed in Stage 2, resulting in N rows per watershed <ref:2606.05265#pg8> How the training signal is built The input data consists of a HEC-RAS simulation database with 592 synthetic storm events across nine Houston-area watersheds <ref:2606.05265#pg4> Each row in the dataset carries 11 predictors and a target, including eight static geospatial features and three dynamic storm predictors <ref:2606.05265#pg4> Events are annotated with two derived fields used for stratification: the most-affected watershed and the RP bin <ref:2606.05265#pg5> The return period is assigned using a procedure that compares peak cumulative rainfall against NOAA Atlas 14 thresholds <ref:2606.05265#pg8> Where the gains come from The selection of in-context examples is central, addressing the issue that naive sampling ignores spatial autocorrelation and event-magnitude imbalance <ref:2606.05265#pg9> The two-stage construction balances event diversity against perwatershed hexagon coverage by setting a maximum context size N = 50k <ref:2606.05265#pg8> Target-aware spatial selection methods, such as FL-Depth, are shown to select hexagons that span the flood response surface by incorporating a z-scored target-depth signal <ref:2606.05265#pg9> Model Performance and Transferability The evaluation shows that Vanilla TabPFN-v2.6 reaches the highest mean R2 of 0.663 across nine watersheds at N = 50k <ref:2606.05265#pg10> This model recovers 98.5% of the Full-KB-XGB reference on held-out watersheds without retraining <ref:2606.05265#pg2> Cross-watershed transfer is achieved via a leave-one-out protocol where the context is built from other watersheds' coresets <ref:2606.05265#pg9> The model's performance on real storms confirms its capability to handle far out-of-distribution cases <ref:2606.05265#pg2> Vanilla TabPFN-v2.6 is identified as the optimal TFM in this evaluation due to its strong pretrained prior <ref:2606.05265#pg10> The dual stratification and target-aware selection are key design elements driving this success <ref:2606.05265#pg10> Future work includes evaluating on a wider set of historical storms and exploring hybrids combining TFM in-context inference with tree-based fitting <ref:2606.05265#pg10> The paper also notes that the 50k-row context cap leaves room to explore larger native windows of TabPFN-v2.6 at roughly 100k <ref:2606.05265#pg10> The mesh-to-H3 conversion standardizes spatial units but future evaluation should report sensitivity at both the H3 and original mesh-cell resolution <ref:2606.05265#pg10> The results here are reported in R2, and complementary metrics such as RMSE, MAE, and high-depth tail accuracy would further characterize operational reliability <ref:2606.05265#pg10> REFERENCES arXiv:2606.05265v1
cs.LG: 3 Jun 2026 Urban Resilience.AI Lab, Zachry Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX, USA bDepartment of Computer Science and Engineering, Texas A&M University, College Station, TX Resilitix Intelligence LLC Institute for a Disaster Resilient Texas, Texas A&M University, College Station, TX A R T I C L E I N F O Keywords: Domain-aware coreset construction In-context learning Tabular foundation model Flood depth prediction Cross-watershed transferability Hydrodynamic surrogate A B S T R A C T Near-real-time flood depth prediction demands surrogate models that are accurate, fast, and transferable across watersheds We propose a domain-aware coreset construction pipeline that conditions a tabular foundation model at inference time The pipeline stratifies storms by return period and most-affected watershed, then samples hexagons with a target-aware spatial selector With 0.7% of the per-watershed training pool, the model attains a mean R2 of 0.663 across nine Houston-area watersheds, within 98.5% of the supervised reference (R2 = 0.673) It transfers to held-out watersheds without task-specific retraining, staying ahead of a coreset-trained supervised baseline On real storms it exceeds the supervised reference on a far out-of-distribution case and trails it on a mostly in-distribution one Domain-aware coreset construction lets tabular foundation models deliver data-efficient, watershed-transferable flood predictions without per-watershed training 1.
Tom: Introduction Timely flood depth information supports emergency managers and infrastructure operators during extreme weather events, and also underpins downstream assessments of flood exposure (Yin and Mostafavi, 2023), mobility disruption, and community resilience (Yin, Li and Mostafavi, 2026) Yet generating it at scale remains computationally demanding (Li, Ma, Brody and Mostafavi, 2025) Physics-based hydrodynamic models such as the Hydrologic Engineering Center’s River Analysis System (HEC-RAS) provide the engineering reference for inundation depth prediction Their cost grows quickly with simulated extent and mesh resolution, especially when many storm events must be screened A direct HEC-RAS sweep is therefore impractical for near-real-time forecasting or broad scenario testing (Ma, Li, Li and Mostafavi, 2026) Machine learning (ML) surrogates trained on simulator output have become a standard alternative and offer near-instant inference once trained (Bentivoglio, Isufi, Jonkman and Taormina, 2022; Mosavi, Ozturk and Chau, 2018) Existing surrogate work falls into two patterns The first trains one model per mesh cell, capturing local dynamics at fine resolution while producing a fragmented collection that cannot be applied beyond the original mesh (Lee, Huang, Antolini, Garcia, Juan, Brody and Mostafavi, 2024) The second trains a single supervised model on the full knowledge base of one watershed, which yields strong inwatershed accuracy but binds the predictor to its training region (Zahura, Goodall, Sadler, Shen, Morsy and Behl, 2020) Both patterns share the same operational drawback: each requires millions of training rows, and adding a new watershed or recalibrating after an event-regime shift entails another full training pass Tree-based gradient boosting (Chen and Guestrin, 2016) still dominates this space because deep tabular methods continue to underperform tree ensembles on structured input (Grinsztajn, Oyallon and Varoquaux, 2022) Deep learning has produced striking results on adjacent hydrologic tasks such as rainfall-runoff modeling (Kratzert, Klotz, Brenner, Schulz and Herrnegger, 2018), but for spatially distributed flood-depth surrogates the treebased pattern remains the practical default Transformer and foundation-model architectures have recently been adapted to structured disaster-management tasks, including post-disaster building-damage categorization (Xiao and Mostafavi, 2025), multimodal impact assessment (Xiao, Yin and Mostafavi, 2026), and graphbased damage prediction (Esparza, Battala and Mostafavi, 2026) Tabular foundation models (TFMs) such as TabPFN (Hollmann et al., 2025) and TabICL (Qu et al., 2024) offer a different path A TFM is a single transformer pretrained on a wide distribution of synthetic tabular tasks It solves a new task at inference time by conditioning on a labeled context set rather than through gradient updates This in-context learning (ICL) framing avoids per-task retraining, and a sufficiently strong pretrained backbone can rival task-specific supervised models on tabular benchmarks However, the context windows of the tabular foundation models considered in this work are bounded at roughly 104 to 105 rows (Hollmann et al., 2025; Grinsztajn, Flöge, Key, Birkel, Jund, Roof, Jäger, Hollmann and Hutter, 2025), while a single watershed’s simulation knowledge base typically reaches millions 3.
Jane: Data The primary data is the MaxFloodCast HEC-RAS simulation database (Lee et al., 2024) covering nine watersheds in Harris County, Texas The database contains 592 synthetic storm events generated by applying a Rasterized Time-series Resampling Method to historic storms in the area, with durations from 1 to 33 hours and hourly rainfall grids at approximately 1,010 m resolution Flood inundation depths are simulated by HEC-RAS 2D over an unstructured mesh of 26,301 cells (nominal size ≈ 366 m), refined along major watercourses and breaklined at high-elevation features From this database we inherit the cell-level peak inundation depth target and the per-event rainfall metadata To establish a uniform spatial unit across all variables, we discretize each mesh cell into H3 Level 10 hexagons (about 75 m edge length) via polygon polyfill (≈ 7.7 hexagons per mesh cell on average) Hexagons in the same mesh cell share that cell’s dynamic features (rainfall, depth) but carry their own static features, yielding 592 events × 177,330 hexagons ≈ 105 million rows with no missing values Each row carries 11 predictors and a target (Table 1) Eight static predictors describe local topography and hydrology: elevation, imperviousness, Topographic Wetness Index (TWI), road density within a 500 m radius, distance to coast, Height above Nearest Drainage (HAND), distance to the nearest stream, and distance to the nearest stream of order ≥ 4 Three dynamic predictors describe each storm: cumulative rainfall, peak rainfall intensity, and precipitation duration Among the static predictors, TWI follows the standard ln(SCA∕ tan β) formulation under D∞ flow routing (Tarboton, 1997) computed in SAGA GIS (Conrad, Bechtel, Bock, Dietrich, Fischer, Gerlitz, Wehberg and Böhner, 2015), and road density is the fraction of road area within a 500 m circular kernel applied to a 15 m-buffered roadway raster The target is the per-event peak inundation depth at each hexagon simulated by HEC-RAS Each storm is annotated with two derived fields used for stratification: the most-affected watershed and the RP bin 4.
Tom: Methodology 4.1.
Paper summary: Tom: So, we're wrapping up our look at this paper, "Data-efficient flood depth prediction through domain-aware coreset selection and tabular foundation models". We've seen how they build a system that conditions a tabular foundation model at inference time using a domain aware coreset.
Jane: The central thesis is that near real time flood depth prediction demands surrogates that are accurate, fast, and transferable across watersheds, and this paper proposes a pipeline to achieve that by compressing the simulation archive first.
Lu: They compress it through a two stage construction: one stage stratifies events by return period and watershed, while the second selects H3 Level ten hexagons with a target aware facility location strategy on page three of this work <ref:2606.05265#pg2>.
Meng: The main claim they make is that vanilla foundation models can match the watershed level baseline with a fifty thousand row coreset, achieving a mean R squared of zero point six six three across nine Houston area watersheds.
Tom: And the key takeaway for us is that this approach allows them to recover ninety eight point five percent of a supervised reference on held out watersheds without retraining, and they transfer cross watershed accuracy using a leave one out protocol on page two of this work <ref:2606.05265#pg2>.
Jane: It seems like they’re showing how you can get decent flood depth predictions just by conditioning the model at inference time instead of doing heavy per watershed fine tuning.
Lalam: The paper shows that this method allows tabular foundation models to deliver data efficient watershed transferable flood predictions without requiring per watershed training.
Conclusion: Tom: Looking at the title, "Data-efficient flood depth prediction through domain aware coreset selection and tabular foundation models", it really captures the essence of what they did. It’s about using smart data sampling to make heavy AI models work on flood prediction tasks without needing massive amounts of per watershed training data.
Jane: I think the implication is that we can move towards faster, more accessible tools for emergency managers because they don't have to wait for huge training runs every time a new area needs predicting.
Lu: This method points toward using foundation models in an in context learning framework, where the coreset acts as the labeled context set to condition the model at inference time.
Meng: From a deployment standpoint, it means less compute per watershed and less maintenance when you have a new area needing coverage. The trade off is that you still need that initial two stage construction pipeline to build that representative subset, which takes upfront effort.
Tom: Exactly. It’s about using targeted data selection—that domain aware coreset construction—to make large tabular models practical for real world application in this kind of hydrology problem.
Lipai Huang, Adithi Srinath, Manas Singh, Junwei Ma, Ali Mostafavi
Urban Resilience.AI Lab · Department of Computer Science and Engineering, Texas A&M University
cs.LG
Submitted: 2026-06-03
Updated: 2026-11-06
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 77/100
The gist: Data-efficient flood depth prediction through domain-aware coreset selection and tabular foundation models is proposed to provide accurate, fast…
Key concepts
- Tabular Foundation Model (TFM)
- A single transformer model pretrained on a wide variety of synthetic tabular tasks. Instead of retraining for every new task, it solves a specific prediction problem by conditioning itself on a small set of labeled examples provided at inference time.
- Domain-Aware Coreset Construction
- A two-stage process to select a compact, hydrologically representative subset of simulation data. Stage 1 stratifies storms by return period and watershed, while Stage 2 uses target-aware spatial selectors to choose hexagons that cover the flood response surface effectively.
- In-Context Learning (ICL)
- The technique where the TFM is conditioned on a small context set of examples during inference. This allows the model to perform a specific task—like predicting flood depth—without needing full retraining, leveraging its strong pretrained knowledge.
Terminology
Summary
Data-efficient flood depth prediction through domain-aware coreset selection and tabular foundation models is proposed to provide accurate, fast, and transferable watershed-level flood depth predictions by conditioning a tabular foundation model at inference time using a domain-aware coreset <ref:2606. p. 1 Data-efficient flood depth prediction through domain-aware coreset
selection and tabular foundation models
Lipai Huanga,∗
, Adithi Srinathb
, Manas Singhb
, Junwei Maa,c and Ali Mostafavia,d
aUrban Resilience. AI Lab, Zachry Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX, USA
bDepartment of Computer Science and Engineering, Texas A&M University, College Station, TX, USA
cResilitix Intelligence LLC, Houston, TX, USA
d
Institute for a Disaster Resilient Texas, Texas A&M University, College Station, TX, USA
A R T I C L E I N F O
Keywords:
Domain-aware coreset construction
In-context learning
Tabular foundation model
Flood depth prediction
Cross-watershed transferability
Hydrodynamic surrogate
A B S T R A C T
Near-real-time flood depth prediction demands surrogate models that are accurate, fast, and transferable across watersheds. The pipeline stratifies storms by return period and most-affected watershed, then
samples hexagons with a target-aware spatial selector. 7% of the per-watershed training pool,
the model attains a mean 𝑅2 of 0. 7% of the per-watershed training pool,
the model attains a mean 𝑅2 of 0. 663 across nine Houston-area watersheds, within 98. p. 1 7% of the per-watershed training pool,
the model attains a mean 𝑅2 of 0. 663 across nine Houston-area watersheds, within 98. The pipeline stratifies storms by return period and most-affected watershed, then
samples hexagons with a target-aware spatial selector. 7% of the per-watershed training pool,
the model attains a mean 𝑅2 of 0. • A demonstration that a vanilla TFM conditioned
on the domain-aware coreset (about 0. 7% of the
per-watershed training pool) recovers 98. 5% of the
supervised reference (𝑅2 = 0. It transfers to held-out watersheds without task-specific retraining,
staying ahead of a coreset-trained supervised baseline.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Domain-Aware Coreset Construction
The pipeline involves a two-stage construction process to compress the physics-based simulation archive into a compact, hydrologically representative subset <ref:2606.05265#pg4> Stage 1 stratifies storm events by return period and most-affected watershed, while Stage 2 selects hexagons with a target-aware spatial selector <ref:2606. p. 2 (2) Domain-aware coreset construction: a two-stage
pipeline compresses the archive into a compact, hydrologically representative subset. Stage 1 stratifies events by return period
and most-affected watershed, and Stage 2 selects H3 Level 10 hexagons with a target-aware facility-location strategy. Stage one stratifies storm
events on two axes: storm return period (RP) and mostaffected watershed. The dual stratification ensures that rare
high-RP events and their host watersheds survive sampling
noise. We address both with a two-stage
construction: events are first dual-stratified by return period
and watershed, then hexagons within each watershed are
sampled with a target-aware spatial selector. The resulting
coreset is the Cartesian product of 𝑁𝑒
events sampled in
Stage 1 and 𝑁ℎ
hexagons selected per watershed in Stage
2, giving 𝑁 = 𝑁𝑒 ×𝑁ℎ
rows per watershed.Where the paper says it
Where the paper says it
Where the paper says it
How the training signal is built
The input data consists of a HEC-RAS simulation database with 592 synthetic storm events across nine Houston-area watersheds <ref:2606.05265#pg4> Each row in the dataset carries 11 predictors and a target, including eight static geospatial features and three dynamic storm predictors <ref:2606. p. 4 Each row carries 11 predictors and a target (Table 1). Eight static predictors describe local topography and hydrology: elevation, imperviousness, Topographic Wetness
Index (TWI), road density within a 500 m radius, distance
to coast, Height above Nearest Drainage (HAND), distance
to the nearest stream, and distance to the nearest stream of
order ≥ 4. Each storm is annotated with two derived fields used for
stratification. The most-affected watershed is the watershed
whose cells have the highest mean inundation depth during
that storm. Category Variable Data Source
Topography
Elevation National Elevation Dataset (NED)
Imperviousness National Land Cover Database (NLCD)
Topographic Wetness Index (TWI) NED + NLCD
Road Density within 500 m radius TxDOT roadway network
Distance to Coast National Hydrography Dataset Plus High Resolution (NHDPlusHR)
Hydrology
Height above Nearest Drainage (HAND) NOAA National Water Model
Distance to Nearest Stream NHDPlusHR
Distance to Stream (order ≥ 4) NHDPlusHR
Aggregated
Event-specific
Cumulative Rainfall HEC-RAS 2D
Peak Rainfall Intensity HEC-RAS 2D
Precipitation Duration HEC-RAS 2D
Event Annotation Return Period (RP) NOAA Atlas 14
event is first assigned a NOAA Atlas 14 duration class 𝑑
∗
based on its storm duration 𝑑𝑒
:
𝑑
∗ =
⎧
⎪
⎨
⎪
⎩
6 h, 𝑑𝑒 ≤ 6 h
12 h, 6 h < 𝑑𝑒 ≤ 12 h
24 h, 𝑑𝑒 > 12 h
(1)
which selects the smallest NOAA Atlas 14 (National Oceanic
and Atmospheric Administration, 2018) class containing 𝑑𝑒
. The event’s peak cumulative rainfall across hexagons, 𝑟max,
is then compared against the Harris-County-median NOAA
Atlas 14 thresholds {𝜏𝑑
∗,𝑇 } at the candidate return periods
𝑇 ∈ {1, 2, 5, 10, 25, 50, 100, 200, 500, 1000} years, and the
event is assigned the largest 𝑇 for which the threshold is
exceeded:
RPevent = max{
𝑇 ∶ 𝑟max ≥ 𝜏𝑑
∗,𝑇 }
, (2)
with RPevent = 0 when no threshold is exceeded.Where the paper says it
Where the paper says it
Where the paper says it
Where the gains come from
The selection of in-context examples is central, addressing the issue that naive sampling ignores spatial autocorrelation and event-magnitude imbalance <ref:2606.05265#pg9> The two-stage construction balances event diversity against perwatershed hexagon coverage by setting a maximum context size N = 50k <ref:2606. p. 5 The decomposition balances event diversity against perwatershed hexagon coverage within a context that fits all
TFMs evaluated. The largest size 𝑁 = 50k matches
TabPFN-v2. The Stage 1 events span below-1-year
through 1,000-year storms, so hexagons selected along
this signal cover the flood-response surface directly. Pure
feature-space selection picks hexagons that look featurediverse but can still cluster on flood response, leaving the
target distribution under-covered.Where the paper says it
Where the paper says it
Model Performance and Transferability
The evaluation shows that Vanilla TabPFN-v2.6 reaches the highest mean R2 of 0. p. 9 rises monotonically with 𝑁 and reaches 𝑅2 = 0. 663 at 𝑁 =
50k, the highest among coreset-based models and a 98. 7% of the per-watershed training pool,
the model attains a mean 𝑅2 of 0. 663 across nine Houston-area watersheds, within 98. • A demonstration that a vanilla TFM conditioned
on the domain-aware coreset (about 0. 7% of the
per-watershed training pool) recovers 98. 5% of the
supervised reference (𝑅2 = 0. It transfers to held-out watersheds without task-specific retraining,
staying ahead of a coreset-trained supervised baseline.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion and Future Directions
Vanilla TabPFN-v2.6 is identified as the optimal TFM in this evaluation due to its strong pretrained prior <ref:2606. p. 9 6 is therefore the strongest
within-watershed model in this evaluation. The stronger
pretrained prior delivers more than per-watershed adaptation
on the v2. For
cross-watershed evaluation we adopt a leave-one-out (LOO)
protocol in which the held-out target watershed contributes
neither training rows nor context. The contributions of this paper are:
• A two-stage domain-aware coreset construction pipeline
that combines event-level stratification by return period and watershed with target-aware hexagon selection, encoding hydrologic and geographic priors in the
context. Its native context
window is much larger than 50k, so the parity cap we apply is
conservative for TabICL and leaves it short of its optimum in
this regime. Vanilla TabPFN-v2. • Models extrapolate on out-of-distribution storms and remain accurate in-distribution. arXiv:2606. Existing surrogate work falls into two patterns. The first
trains one model per mesh cell, capturing local dynamics
at fine resolution while producing a fragmented collection that cannot be applied beyond the original mesh (Lee,
Huang, Antolini, Garcia, Juan, Brody and Mostafavi, 2024).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
REFERENCES
arXiv:2606.05265v1 [cs. p. 0 arXiv:2606. 05265v1 [cs.Where the paper says it
Urban Resilience.AI Lab, Zachry Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX, USA bDepartment of Computer Science and Engineering, Texas A&M University, College Station, TX
Resilitix Intelligence LLC
Institute for a Disaster Resilient Texas, Texas A&M University, College Station, TX
A R T I C L E I N F O
Keywords: Domain-aware coreset construction In-context learning Tabular foundation model Flood depth prediction Cross-watershed transferability Hydrodynamic surrogate
A B S T R A C T Near-real-time flood depth prediction demands surrogate models that are accurate, fast, and transferable across watersheds We propose a domain-aware coreset construction pipeline that conditions a tabular foundation model at inference time The pipeline stratifies storms by return period and most-affected watershed, then samples hexagons with a target-aware spatial selector With 0.7% of the per-watershed training pool, the model attains a mean R2 of 0. p. 1 The pipeline stratifies storms by return period and most-affected watershed, then
samples hexagons with a target-aware spatial selector. 7% of the per-watershed training pool,
the model attains a mean 𝑅2 of 0. 663 across nine Houston-area watersheds, within 98. 5% of the
supervised reference (𝑅2 = 0.Where the paper says it
Where the paper says it
-
Introduction Timely flood depth information supports emergency managers and infrastructure operators during extreme weather events, and also underpins downstream assessments of flood exposure (Yin and Mostafavi, 2023), mobility disruption, and community resilience (Yin, Li and Mostafavi, 2026) Yet generating it at scale remains computationally demanding (Li, Ma, Brody and Mostafavi, 2025) Physics-based hydrodynamic models such as the Hydrologic Engineering Center’s River Analysis System (HEC-RAS) provide the engineering reference for inundation depth prediction Their cost grows quickly with simulated extent and mesh resolution, especially when many storm events must be screened A direct HEC-RAS sweep is therefore impractical for near-real-time forecasting or broad scenario testing (Ma, Li, Li and Mostafavi, 2026) Machine learning (ML) surrogates trained on simulator output have become a standard alternative and offer near-instant inference once trained (Bentivoglio, Isufi, Jonkman and Taormina, 2022; Mosavi, Ozturk and Chau, 2018) Existing surrogate work falls into two patterns The first trains one model per mesh cell, capturing local dynamics at fine resolution while producing a fragmented collection that cannot be applied beyond the original mesh (Lee, Huang, Antolini, Garcia, Juan, Brody and Mostafavi, 2024) The second trains a single supervised model on the full knowledge base of one watershed, which yields strong inwatershed accuracy but binds the predictor to its training region (Zahura, Goodall, Sadler, Shen, Morsy and Behl, 2020) Both patterns share the same operational drawback: each requires millions of training rows, and adding a new watershed or recalibrating after an event-regime shift entails another full training pass Tree-based gradient boosting (Chen and Guestrin, 2016) still dominates this space because deep tabular methods continue to underperform tree ensembles on structured input (Grinsztajn, Oyallon and Varoquaux, 2022) Deep learning has produced striking results on adjacent hydrologic tasks such as rainfall-runoff modeling (Kratzert, Klotz, Brenner, Schulz and Herrnegger, 2018), but for spatially distributed flood-depth surrogates the treebased pattern remains the practical default Transformer and foundation-model architectures have recently been adapted to structured disaster-management tasks, including post-disaster building-damage categorization (Xiao and Mostafavi, 2025), multimodal impact assessment (Xiao, Yin and Mostafavi, 2026), and graphbased damage prediction (Esparza, Battala and Mostafavi, 2026) Tabular foundation models (TFMs) such as TabPFN (Hollmann et al. p. 1
, 2025) and TabICL (Qu et al. p. 1Where the paper says it
Deep learning has produced striking results on adjacent hydrologic tasks such as rainfall-runoff modeling (Kratzert, Klotz, Brenner, Schulz and Herrnegger, 2018), but for spatially distributed flood-depth surrogates the treebased pattern remains the practical default. Transformer and foundation-model architectures have recently been adapted to structured disaster-management tasks, including post-disaster building-damage categorization (Xiao and Mostafavi, 2025), multimodal impact assessment (Xiao, Yin and Mostafavi, 2026), and graphbased damage prediction (Esparza, Battala and Mostafavi, 2026).
, 2024) offer a different path A TFM is a single transformer pretrained on a wide distribution of synthetic tabular tasks It solves a new task at inference time by conditioning on a labeled context set rather than through gradient updates This in-context learning (ICL) framing avoids per-task retraining, and a sufficiently strong pretrained backbone can rival task-specific supervised models on tabular benchmarks However, the context windows of the tabular foundation models considered in this work are bounded at roughly 104 to 105 rows (Hollmann et al. p. 1Where the paper says it
Transformer and foundation-model architectures have recently been adapted to structured disaster-management tasks, including post-disaster building-damage categorization (Xiao and Mostafavi, 2025), multimodal impact assessment (Xiao, Yin and Mostafavi, 2026), and graphbased damage prediction (Esparza, Battala and Mostafavi, 2026). Tabular foundation models (TFMs) such as TabPFN (Hollmann, Müller, Purucker, Krishnakumar, Körfer, Hoo, Schirrmeister and Hutter, 2025) and TabICL (Qu, Holzmüller, Varoquaux and Le Morvan, 2024) offer a different path.
, 2025; Grinsztajn, Flöge, Key, Birkel, Jund, Roof, Jäger, Hollmann and Hutter, 2025), while a single watershed’s simulation knowledge base typically reaches millionsWhere the paper says it
This in-context learning (ICL) framing avoids per-task retraining, and a sufficiently strong pretrained backbone can rival task-specific supervised models on tabular benchmarks. However, the context windows of the tabular foundation models considered in this work are bounded at roughly 104 to 105 rows (Hollmann et al.
-
Data The primary data is the MaxFloodCast HEC-RAS simulation database (Lee et al. p. 4
, 2024) covering nine watersheds in Harris County, Texas The database contains 592 synthetic storm events generated by applying a Rasterized Time-series Resampling Method to historic storms in the area, with durations from 1 to 33 hours and hourly rainfall grids at approximately 1,010 m resolution Flood inundation depths are simulated by HEC-RAS 2D over an unstructured mesh of 26,301 cells (nominal size ≈ 366 m), refined along major watercourses and breaklined at high-elevation features From this database we inherit the cell-level peak inundation depth target and the per-event rainfall metadata To establish a uniform spatial unit across all variables, we discretize each mesh cell into H3 Level 10 hexagons (about 75 m edge length) via polygon polyfill (≈ 7. p. 4Where the paper says it
Data-efficient domain-aware flood depth prediction 3. Data The primary data is the MaxFloodCast HEC-RAS simulation database (Lee et al.
7 hexagons per mesh cell on average) Hexagons in the same mesh cell share that cell’s dynamic features (rainfall, depth) but carry their own static features, yielding 592 events × 177,330 hexagons ≈ 105 million rows with no missing values Each row carries 11 predictors and a target (Table 1) Eight static predictors describe local topography and hydrology: elevation, imperviousness, Topographic Wetness Index (TWI), road density within a 500 m radius, distance to coast, Height above Nearest Drainage (HAND), distance to the nearest stream, and distance to the nearest stream of order ≥ 4 Three dynamic predictors describe each storm: cumulative rainfall, peak rainfall intensity, and precipitation duration Among the static predictors, TWI follows the standard ln(SCA∕ tan β) formulation under D∞ flow routing (Tarboton, 1997) computed in SAGA GIS (Conrad, Bechtel, Bock, Dietrich, Fischer, Gerlitz, Wehberg and Böhner, 2015), and road density is the fraction of road area within a 500 m circular kernel applied to a 15 m-buffered roadway raster The target is the per-event peak inundation depth at each hexagon simulated by HEC-RAS Each storm is annotated with two derived fields used for stratification: the most-affected watershed and the RP binWhere the paper says it
The database contains 592 synthetic storm events generated by applying a Rasterized Time-series Resampling Method to historic storms in the area, with durations from 1 to 33 hours and hourly rainfall grids at approximately 1,010 m resolution. Flood inundation depths are simulated by HEC-RAS 2D over an unstructured mesh of 26,301 cells (nominal size ≈ 366 m), refined along major watercourses and breaklined at high-elevation features.
-
Methodology 4. p. 4
1.Where the paper says it
Imelda mixes indistribution and OOD cells. Methodology 4.
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
-
Data-efficient flood depth prediction via domain-aware coreset construction allows tabular foundation models to deliver data-efficient, watershed-transferable flood predictions without per-watershed training. The pipeline stratifies storms by return period and most affected watershed, then samples hexagons with a target aware spatial selector.
-
In-context learning predicts a held out watershed from neighbors without retraining, enabling the model to
extrapolate on out-of-distribution storms and remain accurate in-distribution.
-
The two stage domain aware coreset construction pipeline combines event level stratification by return period and watershed with target aware hexagon selection, encoding hydrologic and geographic priors in the context. This ensures that the selected context set, which is about 0.7% of the per watershed training pool, captures both event types and watershed geometry.
-
The FL-Depth hexagon selection method is proposed as a default selector for all coreset based models because it has been shown to be
more spread than these spatially neutral baselines because its feature vector includes the z scored depth signal, which varies smoothly in space and therefore makes nearby hexagons redundant under the FL objective.
-
Vanilla TabPFN-v2.6 is identified as the optimal TFM in evaluation because it
reaches mean R2 = 0.663 across nine Houston area watersheds and recovers 98.5% of a Full KB XGB reference trained on roughly 6.9M rows per watershed.
-
The system can perform cross watershed transfer to a held out target watershed by drawing in context examples from its touching neighbors, without any gradient updates, demonstrating that
Vanilla TabPFN-v2.6 stays stable between 0.50 and 0.52 across the full sweep and leads the coresettrained supervised baseline at most context sizes in both source selection modes.
-
The model can outperform a supervised baseline on far out of distribution events like Hurricane Harvey because
the TabPFN pretrained prior generalizes more reliably than tree based fitting under sharp distribution shift, since XGB must extrapolate beyond its training envelope while the TFM stays inside its own prior.
Abstract
Near-real-time flood depth prediction demands surrogate models that are accurate, fast, and transferable across watersheds. Supervised surrogates can match physics-based simulators in accuracy but need millions of training rows per watershed and cannot extrapolate beyond their original mesh. We propose a domain-aware coreset construction pipeline that conditions a tabular foundation model at inference time. The pipeline stratifies storms by return period and most-affected watershed, then samples hexagons with a target-aware spatial selector. With 0.7% of the per-watershed training pool, the model attains a mean R squared of 0.663 across nine Houston-area watersheds, within 98.5% of the supervised reference (R squared = 0.673). It transfers to held-out watersheds without task-specific retraining, staying ahead of a coreset-trained supervised baseline. On real storms it exceeds the supervised reference on a far out-of-distribution case and trails it on a mostly in-distribution one. Domain-aware coreset construction lets tabular foundation models deliver data-efficient, watershed-transferable flood predictions without per-watershed training.
Sources
- Practical Coreset Constructions for Machine Learning
- TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models
- A Parametric Framework for Anticipatory Flashflood Warning: Integrating Landscape Vulnerability with Precipitation Forecasts
- TabICL: A Tabular Foundation Model for In-Context Learning on Large Data
- On Finetuning Tabular Foundation Models
- Unsupervised Graph Deep Learning Reveals Emergent Flood Risk Profile of Urban Areas
Related papers
- A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis