HiRO-ACE: AI Emulation & Downscaling Framework

• HiRO-ACE is a two-stage AI framework that couples ACE2S for stochastic coarse-scale emulation with HiRO for diffusion-based 32x precipitation downscaling.
• It ensures statistical compatibility by training both components on a decade of X-SHiELD data, achieving biases below 10% and accurate reproduction of extreme precipitation up to the 99.99th percentile.
• The framework enables decades of high-resolution regional precipitation simulations in a single day on one H100 GPU, offering a significant energy and time efficiency boost over conventional models.

to=arxiv.search 久赢_code: {"query":"HiRO-ACE fast and skillful AI emulation and downscaling trained on a 3 km global storm-resolving model", "max_results": 5} წლები? to=arxiv.search 天天中彩票中了? Wait previous maybe malformed. Let's do proper. to=arxiv.search ฝ่ายขายรายการ: {"query":"HiRO-ACE fast and skillful AI emulation and downscaling trained on a 3 km global storm-resolving model", "max_results": 5} to=arxiv.search 尚度: {"query":"(Perkins et al., 20 Dec 2025)", "max_results": 5} HiRO-ACE is a two-stage AI modeling framework for emulating and downscaling output from X-SHiELD, a 3 km global storm-resolving atmospheric model. It combines ACE2S, a stochastic version of the Ai2 Climate Emulator running globally at $1^\circ \times 1^\circ$ resolution, with HiRO, a diffusion-based downscaling model that performs 32x downscaling to generate 6-hourly 3 km precipitation over arbitrary non-polar regions of the globe from coarse $\sim 100$ km inputs. Both components are trained on data derived from a decade of X-SHiELD and are designed so that ACE2S outputs can be ingested by HiRO without additional finetuning. In reported evaluations, HiRO-ACE reproduces the distribution of extreme precipitation rates through the 99.99th percentile, achieves time-mean precipitation biases below 10% almost everywhere, and can generate decades of high-resolution regional precipitation within a single day on one H100 GPU (Perkins et al., 20 Dec 2025).

1. System definition and architectural principle

HiRO-ACE is explicitly structured as a coupled emulator-plus-downscaler system. ACE2S first emulates the large-scale atmospheric evolution on a coarse global grid, and HiRO then reconstructs the unresolved kilometer-scale precipitation structure over a selected region. The operational pipeline is: run ACE2S autoregressively to produce global 6-hourly coarse atmospheric fields including surface precipitation and 10 m winds; select a target region and time period; feed the coarse precipitation and wind fields, together with static high-resolution topography, into HiRO; and sample a 3 km precipitation realization over that region (Perkins et al., 20 Dec 2025).

A central design principle is statistical compatibility between the two stages. ACE2S is trained to preserve the coarse-grid precipitation probability distribution and small-scale spectral power of coarsened X-SHiELD, because HiRO is trained in perfect-prediction mode on coarsened X-SHiELD inputs. The paper states that ACE2S “maintains grid-scale precipitation variability consistent with coarsened X-SHiELD, enabling its outputs to be ingested by HiRO without additional tuning,” and reports that no specific HiRO finetuning on ACE2S outputs was necessary (Perkins et al., 20 Dec 2025).

Component	Resolution and cadence	Function
ACE2S	$1^\circ \times 1^\circ$, about 100 km, 6-hourly	Stochastic autoregressive global atmosphere emulator
HiRO	3 km regional, 6-hourly precipitation	Diffusion-based downscaling from coarse precipitation, winds, and topography

Both stages are probabilistic. ACE2S conditions on isotropic Gaussian white noise to sample coarse atmospheric trajectories, while HiRO is a stochastic diffusion model over fine-scale precipitation conditioned on coarse fields and topography. This permits ensemble generation by varying ACE2S noise, HiRO noise, or both, thereby separating uncertainty associated with large-scale atmospheric evolution from uncertainty associated with subgrid precipitation realization (Perkins et al., 20 Dec 2025).

2. Training data, grids, and spatial representation

The training corpus is derived from an 11-year X-SHiELD simulation, with the first year treated as spin-up and the final 10 years retained. The first 9 retained years, 2014–2022, were used for training, and 2023 was held out for independent evaluation. X-SHiELD itself is a global atmosphere simulation at about 3.25 km horizontal resolution on a cubed sphere with 79 vertical levels, forced by historical SST, sea ice, and CO$_2$ (Perkins et al., 20 Dec 2025).

For emulator training, X-SHiELD output was coarsened online to C384, approximately 25 km, then remapped offline to a $1^\circ$ Gaussian grid and vertically coarsened to ACE2S’s 8 vertical layers. For downscaling training, native-resolution 6-hourly precipitation and static surface height were remapped to a custom $5760 \times 11520$ latitude-longitude grid aligned exactly with the $180 \times 360$ Gaussian coarse grid and refined by a factor of 32 in each horizontal dimension within each coarse cell. In this setup, “32x downscaling” means that each $1^\circ$, approximately 100 km, grid cell is refined to a $32 \times 32$ block of approximately 3 km cells (Perkins et al., 20 Dec 2025).

HiRO is trained on paired low- and high-resolution examples extracted from the same X-SHiELD snapshots. Because full global 3 km fields are too large for diffusion training, the data are divided into paired patches of $16 \times 16$ coarse cells and $\sim 100$0 fine cells, corresponding to $\sim 100$1 windows. During training, the starting patch origin is randomly shifted within the global domain and the resulting patch collection is shuffled. At inference time, arbitrary larger regions are tiled with overlapping $\sim 100$2 patches with $\sim 100$3 overlap; generated overlaps are averaged and then cropped, which reduces seam artifacts but introduces some smoothing in the overlap strip (Perkins et al., 20 Dec 2025).

This data representation fixes the scope of the learned target. HiRO-ACE reconstructs 3 km precipitation, not a full multivariate 3 km atmospheric state, and it does so at a 6-hourly cadence. That specialization is fundamental to both the reported efficiency and the reported limitations (Perkins et al., 20 Dec 2025).

3. ACE2S: stochastic coarse atmospheric emulation

ACE2S is a modified ACE2 using a Spherical Fourier Neural Operator architecture with dry-air-mass and moisture conservation constraints inherited from ACE2. Its principal modification is stochastic conditioning: ACE2S replaces ACE2’s instance normalization with conditional layer normalization and conditions on 64 channels of isotropic Gaussian white noise. The stated objective is to avoid collapse to an overly smoothed conditional mean state, because deterministic emulators trained with MSE blur grid-scale precipitation and would therefore drive HiRO off distribution (Perkins et al., 20 Dec 2025).

The model predicts the same 6-hourly atmospheric fields used by ACE2, and the paper explicitly identifies among the outputs required by HiRO the 6-hourly surface precipitation rate and zonal and meridional 10 m wind. Training is two-stage. ACE2S is pretrained on ERA5 for 120 epochs using a one-step loss and then fine-tuned on coarsened X-SHiELD for 120 epochs with multi-step training, ensemble size 2, AdamW at learning rate $\sim 100$4, weight decay 0.01, and EMA decay 0.999. The paper reports that pretraining mattered substantially: training only on the short X-SHiELD record worsened climate biases, including nearly 50% larger precipitation RMSB (Perkins et al., 20 Dec 2025).

The stochastic loss couples probabilistic calibration with spectral realism. The final ACE2S objective is

$\sim 100$5

where the first term is “almost fair” CRPS averaged over grid points and the second is an energy score on spherical harmonic coefficients. In the paper’s interpretation, the afCRPS term encourages calibrated probabilistic predictions at each grid point, while the spectral term penalizes errors in the spatial power spectrum and coherence of the generated fields; the spectral term was reported as essential because nodal CRPS alone produced too much noisy small-scale variability (Perkins et al., 20 Dec 2025).

Model selection was based not only on validation loss but also on climate behavior. During X-SHiELD fine-tuning, the authors ran four-member ensembles over 2014–2018 after each epoch and selected checkpoints by minimizing channel-mean global RMS time-mean bias. Among random seeds, the final model balanced low mean bias and correct precipitation power at the Nyquist scale (Perkins et al., 20 Dec 2025).

Over a 10-year rollout, ACE2S improved substantially on deterministic ACE2 for the variables used by HiRO. The reported global area-weighted RMSBs are 0.18 m/s for eastward 10 m wind, 0.13 m/s for northward 10 m wind, and 0.23 mm/day for surface precipitation, compared with 0.32, 0.20, and 0.48 for ACE2. The paper also states that ACE2S nearly removes the factor-of-three underprediction of precipitation spectral power at the smallest resolved scales that ACE2 exhibited (Perkins et al., 20 Dec 2025).

4. HiRO: diffusion-based 32x precipitation downscaling

HiRO is the second stage of the framework and is derived from CorrDiff. Its target is only 3 km precipitation. The model learns the conditional distribution $\sim 100$6 of high-resolution precipitation $\sim 100$7 given coarse inputs $\sim 100$8, using the residual formulation

$\sim 100$9

where $1^\circ \times 1^\circ$0 is a deterministic baseline and $1^\circ \times 1^\circ$1 is a stochastic residual. Unlike standard CorrDiff, where $1^\circ \times 1^\circ$2 is produced by a learned U-Net regressor, HiRO uses bicubic interpolation of the coarse precipitation field for $1^\circ \times 1^\circ$3, and the diffusion model learns the correction that adds deterministic subgrid effects and stochastic fine-scale structure (Perkins et al., 20 Dec 2025).

HiRO conditions on coarse 6-hourly precipitation, coarse 10 m zonal and meridional winds, and static fine-resolution topography. The topography replaces a positional embedding and is described as important for reproducing terrain-controlled precipitation patterns. The denoiser is a 30M-parameter Song U-Net with 128 channels, 7 levels, one U-Net block per level, base channel embedding multiplier 6, and per-level multipliers $1^\circ \times 1^\circ$4. Training uses standard score-matching and EDM methodology with a log-normal noise distribution $1^\circ \times 1^\circ$5 parameterized by $1^\circ \times 1^\circ$6 and $1^\circ \times 1^\circ$7. The increased $1^\circ \times 1^\circ$8 is stated to ensure enough noise during training to saturate even high-variance residual targets such as tropical cyclones. Inference uses EDM sampling with $1^\circ \times 1^\circ$9 and $_2$0. Training uses batch size 24, Adam, learning rate $_2$1, about 740,000 gradient steps, and EMA decay 0.999 (Perkins et al., 20 Dec 2025).

Checkpoint selection for HiRO is tailored to extremes rather than average predictive error. The paper states that the selected checkpoint minimizes normalized tail bias beyond the validation target 99.99th percentile, evaluated on 76 global snapshots from the held-out year every 10 denoiser epochs. Although the printed equation is malformed in the manuscript, the surrounding text makes the criterion explicit: minimize absolute normalized histogram bias in the extreme tail (Perkins et al., 20 Dec 2025).

In case studies, HiRO reconstructs orographic enhancement in atmospheric rivers, spiral bands and inner-core organization in tropical cyclones, and oceanic or post-frontal convective texture even when those features are absent in the bicubic input. The generated samples are not expected to match the exact target realization, because the conditional problem is one-to-many, but the model is designed to sample plausible realizations from the learned conditional distribution (Perkins et al., 20 Dec 2025).

5. Coupled workflow, evaluation, and reported performance

End-to-end inference is described as follows. ACE2S is initialized from a coarse atmospheric state and run autoregressively at 6-hour cadence for the desired period, producing global $_2$2 fields. A region and time sequence of coarse 6-hourly surface precipitation and 10 m winds are then extracted. For each regional snapshot, those fields together with static 3 km topography are passed to HiRO, which samples one or more 3 km precipitation realizations patchwise and stitches them if necessary (Perkins et al., 20 Dec 2025).

In the near-global band $_2$3S–$_2$4N over the 2023 holdout year, both HiRO perfect prediction and coupled HiRO-ACE are reported to match the X-SHiELD 3 km precipitation histogram well through the 99.99th percentile, around 500 mm/day. In the farthest tail, for precipitation above 1000 mm/day, both underproduce events: the 99.9999th percentiles of HiRO perfect prediction and HiRO-ACE reach 78% and 73% of the X-SHiELD target, respectively. The paper attributes this mainly to rare tropical-ocean convective extremes that appear only as moderate coarse-grid precipitation and are therefore difficult to identify from the 100 km inputs (Perkins et al., 20 Dec 2025).

Time-mean precipitation biases are also reported at regional and global scale. In a western U.S. 2023 perfect-prediction test, HiRO perfect prediction has RMSB 0.2 mm/day and virtually no area-mean bias, versus 0.8 mm/day for bicubic interpolation alone; excluding places with less than 0.5 mm/day climatological precipitation, mean absolute relative bias is about 7.5%. Over the same region in 10-year coupled HiRO-ACE rollouts, the domain-average RMSB is 0.4 mm/day, with most absolute biases under 0.5 mm/day and relative biases under 10%; the largest errors, 1–2 mm/day, occur over high terrain such as the Sierra Nevada and are mostly inherited from ACE2S coarse biases. For global 2023 downscaling, HiRO perfect prediction has RMSB 0.3 mm/day and HiRO-ACE 0.9 mm/day, with negligible global-mean bias (Perkins et al., 20 Dec 2025).

The paper also evaluates event realism and calibration. A landfalling atmospheric river case shows HiRO recovering orographic enhancement and marine frontal texture from a smooth coarse input. Tropical cyclone cases show plausible inner cores, spiral rainbands, and convective organization. In a Philippines tropical cyclone perfect-prediction example, the single X-SHiELD target falls outside the HiRO ensemble 95% confidence interval at only 5% of locations, which the authors interpret as nominal calibration (Perkins et al., 20 Dec 2025).

To assess whether apparent climatological biases exceed internal variability, the paper introduces a noise-floor framework. For an average over $_2$5 ensemble members and $_2$6 years, the post-averaging noise floor is

$_2$7

and a bias is significant if

$_2$8

These formulas are used to argue that many apparent precipitation biases are not statistically distinguishable from internal atmospheric or storm-scale variability (Perkins et al., 20 Dec 2025).

Computational efficiency is a major reported outcome. ACE2S runs at about 1500 simulated years per day on one NVIDIA H100 GPU at $_2$9 resolution. HiRO downscaling of one $1^\circ$0 patch for one simulated year takes about 45 minutes on an H100. The paper states that decades of 6-hourly high-resolution regional precipitation can be generated within a single day on one H100 GPU. Training costs are reported as 7 days on 8 H100s for ACE2S and 4 days on 8 H100s for HiRO. By contrast, the X-SHiELD simulation ran at only 0.12 simulated years per day on 27,648 CPU cores, with total energy use exceeding 800 GJ, compared with less than 10 GJ for training and running the machine-learning models (Perkins et al., 20 Dec 2025).

6. Scope, limitations, and terminological disambiguation

The framework’s scope is explicit. HiRO-ACE currently downscales precipitation only, not a full multivariate atmospheric state. Its target cadence is 6-hourly, so sub-6-hour extremes are not represented. The very far tail of precipitation above the 99.99th percentile remains underestimated, especially isolated tropical convective extremes. The paper also notes slight coarse-cell imprinting in parts of the tropical Pacific and Atlantic ITCZ, oversmoothing in patch-overlap zones, and limited physical consistency across variables because the high-resolution output is precipitation alone. Generalization is fundamentally tied to the training climate and to the available decade of X-SHiELD output; extending to changed climates likely requires new km-scale training simulations (Perkins et al., 20 Dec 2025).

These limitations frame the intended use case. The paper positions HiRO-ACE as an efficient surrogate for long, ensemble, region-specific kilometer-scale precipitation datasets, with intended applications in local climate adaptation planning and extreme event risk assessment. A plausible implication is that the framework is optimized for statistical realism, ensemble generation, and throughput rather than deterministic replay of any single historical event (Perkins et al., 20 Dec 2025).

The name “HiRO-ACE” can create bibliographic ambiguity because closely related acronyms appear in unrelated fields. “HIRO: Hierarchical Information Retrieval Optimization” is a query-time retrieval algorithm for RAG over hierarchical document stores and explicitly does not describe a “HiRO-ACE” variant (Goel et al., 2024). “ACE: Pluggable Adaptive Context Elasticizer across Agents” is an agent context-management module and explicitly does not mention “HiRO-ACE” (Liao et al., 30 Jun 2026). Other unrelated acronym overlaps include “ACE: Adapting sampling for Counterfactual Explanations” (Guerrero et al., 30 Sep 2025), “ACE: Adaptive Constraint-aware Early Stopping in Hyperparameter Optimization” (Chen et al., 2022), “HERO: Heterogeneous Embedded Research Platform for Exploring RISC-V Manycore Accelerators on FPGA” (Kurth et al., 2017), and “HIRO: Heuristics Informed Robot Online Path Planning Using Pre-computed Deterministic Roadmaps” (Huang et al., 2024). In the climate-modeling literature summarized here, HiRO-ACE refers specifically to the coupled ACE2S-plus-HiRO system trained on X-SHiELD output (Perkins et al., 20 Dec 2025).