HyCAM

Hybrid machine-learning & physics atmosphere model

Cloud-resolving physics, learned by neural networks, run in a climate model.

HyCAM replaces the conventional convection and cloud parameterizations of the CAM5 general circulation model with neural networks trained on superparameterized (cloud-resolving) SPCAM simulations. The result is a stable, conservative hybrid model that inherits SPCAM's precipitation physics at a small fraction of its cost, generalizes to a +4K climate, reproduces the MJO in AMIP mode, and works as an ensemble weather-forecast model under data assimilation.

Yilun Han CV (PDF)  ·  Guang Zhang, Principal Investigator

01

Model design

CAM5 host · ResCu + NN-cloud networks · trained on SPCAM

A superparameterized model embeds a cloud-resolving model (CRM) in every grid column; it is physically faithful but roughly two orders of magnitude more expensive than standard physics. HyCAM learns that physics instead: neural networks are trained on SPCAM's grid-scale state and tendencies, then coupled back into the CAM5 host in place of the deep and shallow convection, macrophysics and microphysics schemes.

Schematic: a superparameterized CRM trains an AI parameterization that replaces conventional convection and cloud physics inside the GCM
PDFFig 1.1 · Concept. Grid-scale environments from the GCM dynamical core drive an AI parameterization trained on superparameterization (CRM) output; the network returns rainfall and grid-scale tendencies, replacing the conventional convection and cloud physics.
Physics call sequence of CAM5 versus HyCAM, with the NN module internals: ResCu and NN-cloud networks plus adjustment steps
PDFFig 1.2 · Where the networks sit. Left: the CAM5 physics sequence. Middle: HyCAM swaps the four moist-physics schemes for one NN module; dynamics, radiation, surface coupling, chemistry, PBL and gravity-wave drag are untouched. Right: inside the module, ResCu predicts heating and moistening (dT, dq) followed by supersaturation, precipitation and snowfall adjustments, while NN-cloud predicts the cloud fields (nc, ni, cld) seen by radiation.
02

Climate fidelity

F2000 climatological SST · 10-year free runs · SPCAM as truth, CAM5 as baseline

Three F2000 simulations on the same 1.9×2.5° grid: SPCAM (the cloud-resolving reference), HyCAM, and standard CAM5. HyCAM runs freely for a decade with no drift in global water or energy, and its precipitation climatology sits roughly twice as close to SPCAM as CAM5's in every season.

0.45 vs 0.79
Annual precipitation RMSE vs SPCAM, HyCAM vs CAM5 (mm/day)
0.92 vs 1.50
JJA precipitation RMSE vs SPCAM (mm/day)
0.76 vs 1.07
DJF precipitation RMSE vs SPCAM (mm/day)
25.59 / 25.65 / 25.71
10-yr mean total water, HyCAM / SPCAM / CAM5 (kg/m²): no drift
Ten-year time series of global-mean total precipitable water and total energy for CAM5, SPCAM and HyCAM
Fig 2.1 · Conservation. Global-mean total precipitable water (top) and total energy (bottom) over the free runs. HyCAM tracks the seasonal cycle with stable means (25.59 kg/m², 3.342 GJ/m²) and no long-term drift over 10 years.
Annual-mean surface temperature: SPCAM field, HyCAM minus SPCAM, CAM5 minus SPCAM
Fig 2.2 · Surface temperature. Annual-mean TS. HyCAM's error against SPCAM (RMSE 0.71 K, bias −0.04 K) is on par with CAM5's (0.75 K, −0.30 K), with no large regional artifacts introduced by the network physics.
Precipitation climatology for ANN, JJA and DJF: SPCAM maps and the HyCAM and CAM5 differences from SPCAM
Fig 2.3 · Precipitation, where hybrid physics pays off. Annual, JJA and DJF precipitation. The CAM5 column shows the familiar conventional-physics biases (Indian-monsoon and warm-pool errors, a too-strong double ITCZ); HyCAM roughly halves the RMSE in all three averaging periods.
Zonal-mean temperature and specific humidity: SPCAM fields and the HyCAM and CAM5 differences
Fig 2.4 · Thermodynamic structure. Zonal-mean temperature (top) and specific humidity (bottom). Both models stay within about 1.5 K and 0.16 g/kg of SPCAM; HyCAM is slightly warm aloft where CAM5 is cold. The hybrid physics preserves the mean thermodynamic state while fixing the rainfall.
03

Generalization

+4K uniform SST · out-of-distribution climate · hybrids vs pure ML

The networks never saw a warmer world during training. Run HyCAM in a +4K-SST climate anyway and it reproduces SPCAM's warming response: the upper-tropospheric amplified warming, the moistening pattern, the rain-belt shifts and the intensification of extreme daily rainfall. Pure-ML forecast emulators (ACE2-ERA5) and a coarse hybrid (NeuralGCM) tested the same way lose the response amplitude or the circulation change entirely; a physical host plus learned subgrid physics is what generalizes.

r = 0.94
HyCAM vs SPCAM pattern correlation of the zonal-mean ΔT response
r = 0.93
Same for the zonal-wind response ΔU (NeuralGCM: −0.02)
r = 0.71
ACE2-ERA5 ΔT correlation: pattern survives, amplitude collapses
+4K response of HyCAM and SPCAM: temperature and humidity cross sections, precipitation change map and daily-rainfall PDF
PDFFig 3.1 · HyCAM's +4K response vs SPCAM. Warming and moistening cross sections, the precipitation-change map, and the tropical daily-rainfall distribution, HyCAM (top) against SPCAM (bottom). The out-of-training-range response matches the cloud-resolving reference, including the shift of the extreme-rainfall tail.
Zonal-mean +4K responses of temperature, relative humidity and zonal wind for SPCAM, HyCAM, NeuralGCM and ACE2-ERA5
PDFFig 3.2 · Hybrids vs pure ML under +4K. Zonal-mean ΔT, ΔRH and ΔU responses. HyCAM (second row) tracks the SPCAM truth (r = 0.94 / 0.69 / 0.93). NeuralGCM distorts the humidity and wind responses; ACE2-ERA5, a pure-ML emulator, keeps a washed-out pattern with the amplitude collapsed.
04

AMIP realism

FAMIPC5 1990-2005 · observed SST · evaluated against ERA5 and satellite observations

Driven by observed SSTs for 1990-2005, HyCAM and its CAM5 twin are compared with reanalysis and satellite records. HyCAM reproduces the observed warming trend of land temperature across the AMIP period, and, unlike CAM5, produces a Madden-Julian Oscillation with the observed eastward propagation, spectral peak and lifecycle.

+0.46 / +0.49 / +0.37
1991-2005 land warming trend, ERA5 / HyCAM / CAM5 (°C/decade)
MJO: eastward
HyCAM propagates intraseasonal convection east at the observed speed; CAM5's signal stalls
Yearly land-mean 2-m temperature anomalies 1991-2005 for ERA5, HyCAM and CAM5, relative to each dataset's own period mean, with linear warming trends
Fig 4.1 · Land warming trend. Yearly 2-m temperature anomaly averaged over land, each dataset relative to its own 1991-2005 mean, with linear fits (dotted). All three warm together: ERA5 at +0.46°C/decade, HyCAM at +0.49 and CAM5 at +0.37, and both models follow the observed year-to-year swings (r = 0.87 and 0.90 vs ERA5), from the post-Pinatubo cooling of 1992-93 to the 1998 El Niño peak.
Wavenumber-frequency spectra and lag-longitude correlation diagrams for observations, HyCAM and CAM5
PDFFig 4.2 · MJO spectra and propagation. Wavenumber-frequency spectra and lag-longitude correlations of tropical precipitation and winds. HyCAM concentrates power in the observed eastward wavenumber-1-3, 30-80-day band and propagates coherently across the Maritime Continent; CAM5's variability is weak and quasi-stationary.
Winter OLR MJO lifecycle composites by phase for observations, HyCAM and CAM5
PDFFig 4.3 · MJO lifecycle. Winter OLR composites through the eight MJO phases. HyCAM's convective envelope forms over the Indian Ocean, crosses the Maritime Continent and decays near the dateline as observed; the CAM5 composite stays weak and fragmented.
05

Forecasting with DART

EAKF data assimilation · 40-member ensembles · 14-day forecast of the New Year 1997 California AR

Can a hybrid climate model be a weather model? HyCAM and CAM5 were each cycled through a 40-member ensemble Kalman filter (DART, 36 EAKF cycles, 1996-12-18 to 12-27) and then launched on a single 14-day forecast of the strongest West-Coast atmospheric-river episode of winter 1996/97. Assimilation ends exactly where the forecast begins, so no observations from the verification window leak in. Both hybrids capture the AR landfall days; the animation shows the full two weeks, day by day, against ERA5.

106±20 / 98±16
5-day California precipitation total, HyCAM / CAM5 ensembles (mm)
117
Observed (ERA5) 5-day California total for the same window (mm)
2° and 1°
The whole chain repeated at both resolutions (f19 and f09), both models
GIFFig 5.1 · 14 days, day by day. Daily precipitation (shading) and the IVT = 500 kg m⁻¹ s⁻¹ atmospheric-river contour, 1996-12-27 to 1997-01-09: ERA5 against the four 40-member ensemble means (HyCAM and CAM5 at 2° and 1°). The onset AR of Dec 29-31 is forecast by all four; the California dry-out after Jan 3 is captured; at lead 11 days only the 1° ensembles still hint at the Pacific-Northwest AR.
06

Future directions

Three lines of work now underway

6.1

Physics interpretation

Open the trained networks and read the physics back out: controlled perturbation and swap experiments that map what the learned convection responds to (CAPE, boundary-layer moisture, free-tropospheric humidity), turning the emulator into a probe of convective closure itself.

led by Dr. Guang Zhang
6.2

Weather forecasting

Build out Section 05 into a real forecast system: longer cycled DART assimilation, larger reforecast sets across many events, better ensemble spread, and verification of HyCAM as a subseasonal-to-seasonal prediction model for high-impact rainfall such as atmospheric rivers.

6.3

Merging with JCM, training against ERA5

Port the HyCAM physics into JCM, a differentiable JAX-based global model, so the networks can be trained online, end to end through the dynamics, directly against ERA5 reanalysis instead of SPCAM: the same hybrid design, constrained by the observed atmosphere.