Seismic interferometry provides insights into in situ reservoir response to stress changes.

We present two new seismic velocity models for Alaska from joint inversions of body-wave and ambient-noise-derived surface-wave data, using two different methods. Our work takes advantage of data from many recent temporary seismic networks, including the Incorporated Research Institutions for Seismology Alaska Transportable Array, Southern Alaska Lithosphere and Mantle Observation Network, and onshore stations of the Alaska Amphibious Community Seismic Experiment. The first model primarily covers south-central Alaska and uses body-wave arrival times with Rayleigh-wave group-velocity maps accounting for their period-dependent lateral sensitivity. The second model results from direct inversion of body-wave arrival times and surface-wave phase travel times, and covers the entire state of Alaska. The two models provide 3D compressional- (VP) and shear-wave velocity (VS) information at depths ∼0–100  km. There are many similarities as well as differences between the two models. The first model provides a clear image of the high-velocity subducting plate and the low-velocity mantle wedge, in terms of the seismic velocities and the VP/VS ratio. The statewide model provides clearer images of many features such as sedimentary basins, a high-velocity anomaly in the mantle wedge under the Denali volcanic gap, low VP in the lower crust under Brooks Range, and low velocities at the eastern edge of Yakutat terrane under the Wrangell volcanic field. From simultaneously relocated earthquakes, we also find that the depth to the subducting Pacific plate beneath southern Alaska appears to be deeper than previous models.

The 30 November 2018 magnitude 7.1 Anchorage earthquake occurred as a result of normal faulting within the lithosphere of subducted Yakutat slab. It was followed by a vigorous aftershock sequence with over 10,000 aftershocks reported through the end of July 2019. The Alaska Earthquake Center produced a reviewed aftershock catalog with a magnitude of completeness of 1.3. This well‐recorded dataset provides a rare opportunity to study the relationship between the aftershocks and fault rupture of a major intraslab event. We use tomoDD algorithm to relocate 2038 M≥2 aftershocks with a regional 3D velocity model. The relocated aftershocks extend over a 20 km long zone between 47 and 57 km depth and are primarily confined to a high VP/VS region. Aftershocks form two clusters, a diffuse southern cluster and a steeply west‐dipping northern cluster with a gap in between where maximum slip has been inferred. We compute moment tensors for the Mw>4 aftershocks using a cut‐and‐paste method and careful selection of regional broadband stations. The moment tensor solutions do not exhibit significant variability or systematic differences between the northern and southern clusters and, on average, agree well with the mainshock fault‐plane parameters. We propose that the mainshock rupture initiated in the Yakutat lower crust or uppermost mantle and propagated both upward into the crust to near its top and downward into the mantle. The majority of the aftershocks are confined to the seismically active Yakutat crust and located both on and in the hanging wall of the mainshock fault rupture.

Deep learning recommendation models (DLRMs) have been used across many business-critical services at Meta and are the single largest AI application in terms of infrastructure demand in its data-centers. In this paper, we present Neo, a software-hardware co-designed system for high-performance distributed training of large-scale DLRMs. Neo employs a novel 4D parallelism strategy that combines table-wise, row-wise, column-wise, and data parallelism for training massive embedding operators in DLRMs. In addition, Neo enables extremely high-performance and memoryefficient embedding computations using a variety of critical systems optimizations, including hybrid kernel fusion, software-managed caching, and quality-preserving compression. Finally, Neo is paired with ZionEX , a new hardware platform co-designed with Neo's 4D parallelism for optimizing communications for large-scale DLRM training. Our evaluation on 128 GPUs using 16 ZionEX nodes shows that Neo outperforms existing systems by up to 40× for training 12-trillion-parameter DLRM models deployed in production.

No abstract