<?xml version="1.0"?>

<!DOCTYPE owl [
	<!ENTITY rdf "http://www.w3.org/1999/02/22-rdf-syntax-ns#">
	<!ENTITY rdfs "http://www.w3.org/2000/01/rdf-schema#">
	<!ENTITY xsd "http://www.w3.org/2001/XMLSchema#">
	<!ENTITY owl "http://www.w3.org/2002/07/owl#">
	<!ENTITY cc "http://web.resource.org/cc/#">
	<!ENTITY project "http://ebiquity.umbc.edu/ontology/project.owl#">
	<!ENTITY person "http://ebiquity.umbc.edu/ontology/person.owl#">
	<!ENTITY pub "http://ebiquity.umbc.edu/ontology/publication.owl#">
	<!ENTITY assert "http://ebiquity.umbc.edu/ontology/assertion.owl#">
]>

<!--

This ontology document is licensed under the Creative Commons
Attribution License. To view a copy of this license, visit
http://creativecommons.org/licenses/by/2.0/ or send a letter to
Creative Commons, 559 Nathan Abbott Way, Stanford, California
94305, USA.

-->

<rdf:RDF 
		xmlns:rdf = "&rdf;"
		xmlns:rdfs = "&rdfs;"
		xmlns:xsd = "&xsd;"
		xmlns:owl = "&owl;"
		xmlns:cc = "&cc;"
		xmlns:project = "&project;"
		xmlns:person = "&person;"
		xmlns:pub = "&pub;"
		xmlns:assert = "&assert;">
	<pub:TechReport rdf:about="http://ebiquity.umbc.edu/paper/html/id/1230/A-Path-to-Sub-Km-Regional-Emulation-with-the-NOAA-HRRRCast-Lateral-Height-Dependent-Boundary-Forcing-and-a-Resolution-Invariant-Nesting-Methodology">
		<rdfs:label><![CDATA[A Path to Sub-Km Regional Emulation with the NOAA HRRRCast: Lateral Height-Dependent Boundary Forcing and a Resolution- Invariant Nesting Methodology]]></rdfs:label>
		<pub:title><![CDATA[A Path to Sub-Km Regional Emulation with the NOAA HRRRCast: Lateral Height-Dependent Boundary Forcing and a Resolution- Invariant Nesting Methodology]]></pub:title>
		<pub:publishedOn rdf:datatype="&xsd;dateTime">2026-11-03T00:00:00-05:00</pub:publishedOn>
		<pub:abstract><![CDATA[Two non- retraining methods move HRRRCast toward nested sub-kilometer forecasting, (i) Height-
dependent lateral relaxation toward GFS cuts 48-hour forecast errors of the 6 km HRRRCast diffusion
model by 14–43% in a summer and a winter case, and (ii) coupling a nest to a CONUS HRRRCast host,
with the same relaxation at the nest's edges, lets the same 6 km model run on 3 km sub-domains with
lower error than its own CONUS forecast over the same area.
<br>HRRRCast advances its state at a 6-hour cadence and sees GFS only as a full-domain conditioning at four
pressure levels. Its errors grow fastest near the domain edges: in summer along the northern boundary,
in winter across the whole domain and most at the upwind (western) edge. We relax the model state
toward regridded GFS in edge zones whose width at each level equals the distance air travels in 6 hours
at a climatological wind speed (22 cells at 1000 hPa to 180 cells at 200 hPa on the 6 km grid). The zones
are applied at every 6-hour boundary, on all 20 levels and the 10 m and 80 m winds.
<br>At 48 hours, a whole-domain RMSE is reduced 29–39% for 500 hPa height and 14–20% for mean sea-
level pressure, which is never relaxed directly. Edge-band errors are reduced 60–82%; interior gains are
smaller (3–4% in summer, 7–14% in winter). Relaxing only the four GFS conditioning levels makes every
unrelaxed level worse (up to 2.6× baseline in edge bands). Relaxing only at inflow points, the physics-
model convention, removes most of the benefit: the network's edge errors arise at outflow edges too.
<br>We also find that HRRRCast's GFS inputs pair earth-relative GFS winds with grid-relative HRRR winds on
the Lambert grid, a rotation of up to 23° at the east and west edges. This is still present in the current
public code. The network appears to have learned to compensate, which hides the error from standard
verification.
<br>Run on native 3 km input over four nested sub-domains, the 6 km HRRRCast matches its own 6 km skill
on the same domain (error ratios 0.7–1.2 to 48 hours), supporting near-resolution invariance. Small
stand-alone nests whose western edge lies over land collapse within one 6-hour step at either
resolution. We trace this to the normalization of the nested OROG network's global max pooling: taking
the max-pooled values of a full-CONUS HRRRCast, run in parallel with the nested HRRRCast run, makes
the NE nest's 6-hour 500 hPa height error reduce from 268 m to 19 m, matching the CONUS forecast.
This coupled CONUS HRRRCast -in-HRRRCast nesting. through these max pools at every 6-hour step,
avoids retraining the model. The NE and SE nests at both 3 km and 6 km stay near CONUS skill to 48 hours (F48 500 hPa height error 0.90–1.28 of the CONUS run, against 2.4–3.1× for the stand-alone
nests), with 3 km and 6 km nests within 0.96–1.28 of each other. Adding the GFS relaxation zones at the
coupled nests' edges lowers the F48 500 hPa height error to 0.54–0.86 of the CONUS run (3 km nests;
0.58–0.90 at 6 km). Relaxing the edges toward the CONUS HRRRCast host instead gains only about 10%,
since it adds no information the host lacks.
<br>These results outline a path to 1 km nests built from GSL's native 3 km HRRRCast, by the same steps:
initialize the nest from a 1 km analysis, normalize static fields with the model's fixed training statistics,
couple the nest to a CONUS host through the global max pools (or retrain on random sub-domain crops),
and relax toward the driving model in zones sized to 6-hour advection. Resolution invariance from 3 km
to 1 km is the hypothesis we test next.]]></pub:abstract>
		<pub:counter>61</pub:counter>
		<pub:tag><![CDATA[ai weather prediction]]></pub:tag>
		<pub:tag><![CDATA[channel attention (cbam)]]></pub:tag>
		<pub:tag><![CDATA[diffusion models]]></pub:tag>
		<pub:tag><![CDATA[hrrrcast]]></pub:tag>
		<pub:tag><![CDATA[lateral boundary conditions]]></pub:tag>
		<pub:tag><![CDATA[limited-area modeling]]></pub:tag>
		<pub:tag><![CDATA[nesting]]></pub:tag>
		<pub:tag><![CDATA[resolution invariance]]></pub:tag>
		<pub:booktitle><![CDATA[CARTA UMBC Tech Report]]></pub:booktitle>
		<pub:author>
			<rdf:List>
				<rdf:first>
					<person:Person rdf:about="http://ebiquity.umbc.edu/person/html/Milton/Halem"><person:name><![CDATA[Milton Halem]]></person:name><rdfs:label><![CDATA[Milton Halem]]></rdfs:label></person:Person>
				</rdf:first>
				<rdf:rest>					<rdf:List>
						<rdf:first>
							<person:Person rdf:about="http://ebiquity.umbc.edu/person/html/Fnu/Sirisha"><person:name><![CDATA[Fnu Sirisha]]></person:name><rdfs:label><![CDATA[Fnu Sirisha]]></rdfs:label></person:Person>
						</rdf:first>
						<rdf:rest rdf:resource="&rdf;nil" />
					</rdf:List>
				</rdf:rest>
			</rdf:List>
		</pub:author>
		<pub:firstAuthor>
<person:Person rdf:about="http://ebiquity.umbc.edu/person/html/Milton/Halem"><person:name><![CDATA[Milton Halem]]></person:name><rdfs:label><![CDATA[Milton Halem]]></rdfs:label></person:Person>
		</pub:firstAuthor>
		<pub:softCopy><pub:SoftCopy>
			<pub:softCopyFormat><![CDATA[PDF Document]]></pub:softCopyFormat>
			<pub:softCopyURI><![CDATA[http://ebiquity.umbc.edu/get/a/publication/1520.pdf]]></pub:softCopyURI>
			<pub:softCopySize>553948</pub:softCopySize>
			</pub:SoftCopy>
			</pub:softCopy>
	</pub:TechReport>

<rdf:Description rdf:about="">
	<cc:License rdf:resource="http://creativecommons.org/licenses/by/2.0/" />
</rdf:Description>

</rdf:RDF>
