With 19 channels across four instrument feedhorns, AWS spans three key absorption bands for atmospheric sounding. These include temperature sounding near 50 GHz and humidity sounding near the 183 and 325 GHz water vapour absorption lines. Most of these channels are for atmospheric sounding, but there are also three window channels at 50.3, 89.0, and 165.5 GHz to aid in retrieval of surface properties. This channel complement (see Table ) is similar to that of heritage instruments like the Advanced Technology Microwave Sounder (ATMS; ) or the combination of AMSU-A and MHS, but without lower-frequency window channels at 23 and 31 GHz and without channels sounding the mid and upper stratosphere.

Some key differences between AWS and heritage MW sounders are worth noting. First, AWS holds the first channels above 300 GHz – sub-millimetre wavelengths and thus “sub-mm” channels – on an instrument intended for near-real time, operational use ; these channels provide unprecedented sensitivity to atmospheric ice mass and presage the capabilities of the next-generation Ice Cloud Imager (ICI) from EUMETSAT , expected to launch in 2026. Second, to enable a compact design and save weight, the AWS radiometer does not use a quasi-optical network with beam co-alignment, which many heritage instruments rely upon to co-locate all channels . This means that the four feedhorns of AWS point in slightly different directions, leading to disparate geolocations, incidence angles, and swaths on the ground between channel sets. Third, AWS 183 GHz channels feature single-sidebands, whereas on heritage sounders these channels typically feature dual-sidebands. This instrument design choice does not significantly impact the geophysical characteristics, and herein we consider them equivalent for sake of comparison, e.g. 183.31 ± 3 GHz on ATMS and 180.31 GHz on AWS.

AWS was designed to fly in a variety of sun-synchronous and low-inclination orbits, as the specifications for the EPS-Sterna constellation had not yet been determined. From the ride-share launch in August 2024, AWS was inserted into a 22:30 local time (LT) of the ascending node (LTAN) sun-synchronous orbit. This puts it slightly after the Metop series at 21:30 LT and before the operational NOAA satellites at 01:30 LT on their descending node. During the first months of AWS there were three legacy NOAA satellites in drifting sun-synchronous orbits (NOAA-15, -18, -19), but these were decommissioned in June 2025. Figure  demonstrates the importance of orbital crossing times for sounder data, putting AWS observations in the context of cross-track sounder radiances used operationally at ECMWF in a given 90-minute window in late 2025.

AWS orbits at approximately 600 km altitude, lower than the Metop and NOAA platforms at about 830 km. This lower altitude allows for relatively small field of view (FOV) sizes on the ground, with 10 km FOVs achieved at the higher frequencies from an antenna with effective diameter of only 0.16 m. Linked to its lower orbiting altitude, to achieve a roughly 2000 km swath width from all feedhorns, AWS observes out to scan angles of 54.5°, greater than the 52.725 of ATMS or 49.44 of MHS, for example. This leads to larger zenith angles that can reach 70° in some cases.

The AWS scan pattern features higher incidence angles at swath edge and greater spatial oversampling along- and across-track than most heritage instruments. AWS rotates at 50.4 rpm, producing 145 Earth-view samples per scan line in 110° of scan angle from a 2.5 ms integration time, and the near-nadir cross-track samples are separated by about 8 km. We can contrast this with ATMS, which samples with 16 km scan spacing and 90 spots per scan. An illustration of the differing swaths and the FOV variation across the scan can be seen in Fig. . This figure also displays the different geolocations of the four feedhorns.

Due to the different geolocations of the feedhorns on AWS, a method is needed to bring the observations together into a single observation vector. Although this is not strictly necessary, it is what data assimilation systems and radiative transfer models are built to handle, and it provides certain scientific synergies such as allowing lower frequencies to retrieve the surface emissivity for higher frequencies that have less surface sensitivity.

The problem of co-locating the AWS fields of view is solved by superobbing onto a common equal-area grid of around 50 km. Superobbing is a method to spatially average the original observations , with the feedhorn geolocations determining which box a given observation is assigned to. It is a simple average, with antenna response patterns disregarded. Once each feedhorn has been averaged separately, we can build 19-channel observation vectors at a single geolocation due to the exact spatial matches, i.e. match the horns together. At swath edges where not all horns overlap, superobs may be populated by one or two horns only (for example over western Sicily in Fig. ). Zenith angle information is retained on a per-feedhorn basis, which is necessary for realistic simulation of the radiances. The superobbing and horn-matching method simplifies downstream processing, reduces data volume considerably, and importantly reduces representation errors for later assimilation. For example, 35 AWS observations can be averaged together into one 50 km superob in the middle of the swath; this also has a significant impact on effective instrument noise, as we will see later.

Superobbing AWS radiances to 50 km resolution follows the method applied to other humidity sounders MHS and MWHS-2 for assimilation in the IFS . In contrast, the temperature sounder AMSU-A has much larger footprints and thus its observations are not superobbed. For ATMS, it is currently assimilated in clear-sky conditions only after 3 × 3 superobbing , and remains assimilated by this method in experimentation herein. However, for comparisons of radiometric behaviour in this paper, ATMS is treated as AWS, i.e. 50 km superobs in all-sky conditions.

This section briefly describes newer developments of all-sky assimilation at ECMWF and those relevant for the remainder of the paper. Fuller descriptions of the all-sky assimilation method can be found elsewhere in the literature e.g, as well as in the official IFS documentation .

All-sky radiance assimilation is underpinned by a radiative transfer model that simulates the effects of clouds and precipitation in addition to atmospheric gases. In this study we use RTTOV-SCATT version 13.2, which importantly includes updated scattering properties of frozen hydrometeors and the SURFEM-Ocean emissivity model over sea . Both of these developments were done with the advent of sub-mm radiances in mind, and are included in the IFS version used in this study, Cycle 49r1 . In addition, RTTOV-SCATT takes as input ozone profiles from the IFS (rather than climatological profiles), which is important for AWS due to non-negligible ozone sensitivity for sub-mm frequencies . Lastly for radiative transfer, RTTOV uses the measured spectral response functions (SRFs) from laboratory measurements of AWS to compute absorption coefficients.

To assimilate radiances from cloudy and precipitating scenes, the IFS uses a symmetric observation error model that equally weights observed and modelled cloud amounts when determining the observation error. This relies on a cloudiness proxy to indicate clear-sky and cloudy scenes, as representation error is the main driver of all-sky observation errors and this is highly correlated with cloudiness. Whereas heritage sounders like AMSU-A and MHS use window channel differences for cloud proxies , AWS lacks the low-frequency channels used by AMSU-A and also has a new sub-mm channel set that may need a different approach. Thus a useful indicator of cloudiness for AWS is the cloud impact (CI; see Eq. ), which has been used previously in the infrared . CI is defined as 1CI=(O-BiasCorr)-Bclr/2+B-Bclr/2 with B denoting the short-range forecast in observation space (i.e. model background), simulated for both all-sky conditions and in clear-sky (Bclr). A bias correction term (BiasCorr) is included from the variational bias correction (VarBC) scheme to remove systematic offsets between observations (O) and model.

In general, the assimilation of AWS radiances in the IFS follows the methods in place for AMSU-A and humidity sounders MHS and MWHS-2 . However, three aspects of AWS require additional methodological consideration: its lack of low-frequency channels for liquid water path (LWP) retrieval, its novel 325 GHz channels, and the dependency of effective instrument noise on scan position due to superobbing. These are covered here in order.

As defined above, CI is a relatively simple way to define cloudiness in terms of brightness temperatures (TBs). As investigated by , the lack of 23.8 and 31.4 GHz low-frequency window channels relative to AMSU-A does not hinder assimilation of the 50 GHz sounding band in all-sky conditions. Instead of LWP, we use the cloud impact at AWS channel 2 (CI₂; equivalent to AMSU-A channel 4) to guide the observation error model for channels 4 to 7. In Fig. , the left panels show the observation error models for the AWS temperature sounding channels on horn 1 as a function of CI₂. AWS uses CI over both land and sea, in contrast to AMSU-A, which uses LWP over sea and scattering index (SI) over land. Following the all-sky implementation of instruments like AMSU-A, the observation error models shown have been determined empirically by the behaviour of O–B standard deviations as a function of the symmetric cloud proxy. For 183 GHz channels, the SI approach of is retained, using a difference of AWS channels at 89 and 165.5 GHz (see middle panels in Fig. ).

The novel 325 GHz humidity sounding channels on AWS have greater sensitivity to ice particle scattering than 183 GHz channels, and it follows that the cloud proxies previously used at 183 GHz may not be sufficient at sub-mm frequencies. analysed this in detail and determined that a CI-based approach was preferred, using the lowest-peaking 325 GHz channel (19 on AWS). CI₁₉ is thus used for the observation error model of 325 GHz channels over both land and sea, with values seen in the right panels of Fig . In terms of quality control decisions, 325 GHz channels follow exactly what is done for 183 GHz channels with equivalent sensitivity (e.g. channel 18 is treated the same way as channel 13; see ).

Superobbing decreases effective instrument noise in a way that is significant for temperature sounding channels but largely negligible for humidity-sensitive frequencies . This is because background errors for temperature sounding channels are an order of magnitude smaller than those of humidity sounding and window channels, with sample NEDT typically larger than background errors. When employing grid-based superobbing, cross-track sounders like AWS achieve greater noise reduction where more observations are averaged together. Near the middle of the swath roughly 30 observations are averaged together and thus noise is greatly reduced near nadir, compared to little noise reduction at scan edge where perhaps only 2 or 3 observations constitute one superob.

With effective instrument noise a key driver of observation errors, it becomes necessary to augment the all-sky observation error model for superobbed temperature sounding channels on cross-track instruments. Here we use zenith angle as a proxy for the number of observations per superob, and define a noise term (σzen) that is scaled as a function of the zenith angle in radians (θ), given in Eq. (). This term is added in quadrature to the total observation error as output by the standard all-sky error model in clear or cloudy conditions (σclr+cld), seen below in Eq. (). 2fscale(θ)=0.1+0.9exp⁡(-0.8θ2)3σtotal=σclr+cld2+σzen/f(θ)2

This error model augmentation is applied to AWS channels 4 to 7, with σzen ranging from 0.10 to 0.14 (see Table 11 in ). The values for σzen and the functional form for fscale were determined empirically from statistics of O–B standard deviations as a function of scan position, endeavouring to flatten the curve and improve Gaussianity of normalised departures overall. As formulated, the additional term causes no inflation at nadir and produces enough inflation at scan edges to account for greater effective noise in superobs with fewer observations.

Short-range forecast impact can be judged by changes in background fits to independent observations. Figure  shows the normalised change in SD(O-B) for six different observation types spanning temperature, humidity, and wind measurements. The 100 % line represents the control experiment (i.e. without AWS assimilated), so points below this indicate an improved fit between the observations and the model background after adding AWS to the assimilation, i.e. an improved short-range forecast. AWS clearly improves the short-range forecasts of humidity, with decreased SD(O-B) seen for ATMS humidity channels 18–22, several of IASI's humidity-sensitive wavenumbers of 1367 to 1994, and water vapour channels from GOES-18 and Meteosat-9, though radiosonde humidity fits appear largely neutral. There are also some small but significant improvements in fits to wind observations in the upper troposphere, seen in panel (c). The impact on short-range temperature forecasts is relatively small, with an overall neutral change seen in hyperspectral infrared instruments like IASI but some slight improvements seen for radiosonde temperature in the mid troposphere and ATMS channels 6–9.

Not every signal in Fig.  is positive, but this picture is generally consistent with activations of MW sounders from heritage platforms in the IFS e.g.. Due to the large number of humidity-sensitive channels on AWS (five at 183 GHz plus three active at 325 GHz), it is not surprising that the main impact from AWS is on humidity. An interesting feature here is that fits to GOES-18 and Meteosat-9 are quite significantly improved but impacts are neutral for the others in the “geo ring”. This is linked to strong improvements in background fits to observations in two distinct longitude bands near 45° E and 135° W. This appears to be linked to the orbital crossing time of AWS (10:30 LT), filling in a gap between other sounder overpasses (see Fig. ). AWS provides extra constraint for humidity late in the 12 h assimilation window for these longitude bands, which has a larger impact on the forecast in a well-documented effect . It is a timeslot that otherwise has limited MW data at these longitude bands (being about 1 h after Metop-B and -C), and this manifests in improved fits to GOES-18 and Metosat-9 infrared radiances once AWS is added.

To analyse the medium-range impact we examine analysis-based forecast verification. Statistical significance testing of medium-range forecast impacts follow , with 95 % confidence levels given. Forecast impacts are gauged by forecast error reduction in reference to the control, with negative values signifying less error in the experiment and thus a positive impact on the forecast from assimilation of AWS.

Figure  presents the changes in RMSE for winds at two levels (850 and 200 hPa) and geopotential height at 500 hPa, spanning forecast lead times of 2 to 7 d. Significant improvements in forecast skill can be seen in the Southern Hemisphere through day 4, with relatively neutral impacts seen in the tropics and Northern Hemisphere. This is roughly in line with the impacts seen from adding another MWHS-2 into the IFS , but with a slightly larger impact on temperature. This indicates that AWS is able to positively affect medium-range forecast skill in a similar way as heritage MW sounders assimilated in the IFS.

To visualise the spatial distribution of these impacts, Fig.  shows the normalised change in RMSE for winds at four forecast lead times. These plots indicate that the majority of the forecast improvements from AWS are seen at high latitudes in both hemispheres. From the free troposphere down to the surface, winds are improved by about 1 % to 3 % at short-range near the poles and some of this signal lasts through to day 4 forecasts. Similar signals are witnessed in verification of humidity forecasts (not shown).

Another way to characterise the influence of assimilated observations in NWP is through the adjoint-based Forecast Sensitivity to Observation Impact (FSOI). The FSOI metric estimates whether each observation reduced or increased forecast errors at a 24 h lead time based on analysis verification with respect to a dry total energy norm . Here we will use FSOI from the operational system, which uses 12 h assimilation windows as before but runs at 9 km horizontal resolution with a final inner-loop resolution of 40 km. AWS has been assimilated operationally since 10 July 2025, and the statistics here run from this date through October 2025. Note that FSOI statistics give only broad estimates of short-range forecast impacts of contributing observations in the presence of the full observing system used, and limitations of FSOI and its complementarity to OSEs are further discussed by .

The geographic distribution of total FSOI for AWS channel groups 4–7, 11–15, and 16–18 is shown in Fig. . As discussed above, the main impact of AWS assimilation is felt in distinct longitude bands that correspond to the last hours of the 12 h assimilation window observed by AWS with its 10:30 orbital crossing time (see Fig. ). In terms of relative magnitudes, the 183 GHz channels are the most impactful set, followed by the 50 GHz temperature sounding channels. The 325 GHz channels do provide positive impact, most notably in regions near Antarctica, but less in aggregate than the 183 GHz suite. Though FSOI can be a noisy diagnostic especially on short timescales, it is interesting that some regions exhibit what appears to be sustained negative impact. For instance, temperature sounding channels off the Horn of Africa and the 183 GHz suite east of Asia show some small but non-negligible degradation in FSOI. These seem to coincide with areas of known SST biases, and warrant further investigation.

Alongside the total FSOI maps for AWS channel groups, Fig.  also shows MWHS-2 from FY-3E. As FY-3E flies in a sparsely sampled orbit (05:30 or “early morning”), and MWHS-2 has five equivalent 183 GHz channels, it provides a useful point of comparison. Following AWS, the channels of MWHS-2 are grouped, now with 118 GHz temperature sounding channels (2–7) and 183 GHz humidity channels (11–15) shown together. The total FSOI patterns for FY-3E MWHS-2 are remarkably similar to AWS for the 183 GHz suite, albeit shifted as expected from the different equator crossing time, with large impacts in the Southern Ocean especially. FY-3E also shows the clear signal of impacts maximising in two longitude bands, again linked with timeslots late in the 12 h assimilation windows (compare with Fig. ). The FSOI results thus back up the background fits to geostationary infrared radiances (Fig. ), linking impact to orbital crossing time. Lastly, if we compare the impact for MWHS-2 118 GHz channels with those of AWS 50 GHz channels, it is clear that the 50 GHz suite is more impactful. This is especially the case in the persistently cloudy Southern Ocean, where the 50 GHz suite penetrates clouds better than at 118 GHz, which is much more sensitive to attenuation from cloud and precipitation. In addition, the 50 GHz suite typically displays lower sensor noise than 118 GHz, which is crucial for temperature increments.

In late 2025, AWS and FY-3E MWHS-2 are typically two of the most impactful satellite instruments assimilated in the IFS as judged by FSOI. The large FSOI totals seem largely driven by the unique orbital crossing times of these two satellites. It is also clear that AWS achieves significant impact from each suite of sounding channels, though the sub-mm suite is currently the least impactful of the three. As this is the first time that sub-mm channels are assimilated in global NWP, there is surely scope for refinement, and these achieved contributions are already very encouraging.