Supercomputing Frontiers and Innovations

Distillation for Adaptation Language Models to Russian Language

Grigory P. Kovalev, Mikhail M. Tikhomirov — Wed, 21 Jan 2026 00:00:00 +0500

Adapting large language models (LLMs) to morphologically rich languages like Russian presents a major challenge, as multilingual models often exhibit limited transfer due to predominantly English-centric pre-training. This study investigates knowledge distillation (KD) as a more effective alternative to supervised fine-tuning (SFT) for the final calibration stage of language adaptation. We introduce an efficient offline top-K distillation approach that transfers knowledge from a 32B Russian-adapted teacher model to a 4B student model through tokenizer alignment and direct logit transfer. Experimental results demonstrate that KD consistently surpasses SFT, achieving up to a 4.22% performance improvement, with top-100 distillation yielding the highest gains (3.27% on average) albeit with increased memory consumption (62 GB vs. 7 GB for top-10).Moreover, the advantages of KD are most pronounced for student models with lower adaptive capacity (i.e., smaller LoRA α values). These findings underscore the efficacy of KD as a practical and scalable approach for language adaptation, while emphasizing the necessity of balancing performance improvements against computational efficiency.

Supercomputing Co-Design for Solving Ill-Posed Linear Inverse Problems Using Iterative Algorithms

Alexander S. Antonov, Vladimir V. Voevodin, Dmitry V. Lukyanenko — Wed, 21 Jan 2026 00:00:00 +0500

The paper considers an approach to applying the ideas of supercomputing co-design for the effective use of arbitrary multiprocessor computing systems with distributed memory when using iterative regularization algorithms to solve ill-posed linear inverse problems, which are reduced to solving large overdetermined systems of linear algebraic equations with a dense matrix. The proposed methodology allows for a large number of algorithms to select the best virtual topology of processes (in terms of parallelization efficiency) for solving problems of the class under consideration within the allocated resources of the supercomputer system being used.

Supercomputer Methods of Ultrasound Tomographic Imaging in NDT Based on Lamb Waves

Wed, 21 Jan 2026 00:00:00 +0500

This article concerns the developing of supercomputer methods for solving inverse problems of ultrasonic tomography in application to nondestructive testing of thin plates using Lamb waves. Such problems are computationally expensive, as longitudinal, shear, and other waves propagate in solids, requiring the use of vector elastic models of wave propagation. Iterative methods for solving the inverse problem have been developed. The methods are based on gradient descent methods of minimizing the residual functional. Efficiency of the proposed algorithms is illustrated on model problems. The field of wave tomography, which is currently under development, requires powerful computing resources. Parallel computations in this study have been performed on generalpurpose processors of the Lomonosov supercomputer complex. Solving the Helmholtz equation is the core element of the developed algorithms for solving inverse problems of wave tomography. The most demanding computations involve solving linear equations with large-scale sparse matrices using LU-decomposition method. The algorithms were implemented using linear algebra libraries with serial and parallel code. Effectiveness, scalability and performance of the method has been evaluated on CPU computing platforms.

Efficient Algorithm Based on the Woodbury Formula for Modeling Multi-port Antenna Systems

Rufina M. Tretiakova, Alexey V. Setukha, Ilya A. Mass — Wed, 21 Jan 2026 00:00:00 +0500

This work presents an efficient computational approach for modeling antenna systems with multiple ports using boundary integral equations. The method employs the RWG basis functions within the Galerkin scheme to evaluate for surface currents. A key challenge addressed is the repeated solution of linear systems when calculating mutual coupling characteristics (impedance matrix, S-parameters, VSWR) for various port loading conditions. To overcome this, an algorithm based on the Woodbury formula is developed, significantly reducing computational costs by leveraging the low-rank nature of port-related matrix modifications. The method's effectiveness is demonstrated for both wire and patch antenna arrays, showing substantial speedups—approximately proportional to the number of ports for direct solvers and significant gains for iterative solvers using mosaic-skeleton approximations while maintaining solution accuracy.

GPU-based Large-eddy Simulation of Mixed-phased Clouds

Evgeny V. Mortikov, Elizaveta M. Gashchuk, Andrey V. Debolskiy — Wed, 21 Jan 2026 00:00:00 +0500

We discuss the development of the large-eddy simulation (LES) model of the atmospheric boundary layer with embedded two-moment bulk cloud microphysics scheme well-suited for massively-parallel heterogeneous supercomputers based on GPU (Graphics Processing Units) architecture. To evaluate the LES model and its computational efficiency, we consider the numerical setup corresponding to the development of an intense Arctic cold-air outbreak case. It is shown that the dynamic closure approach for calculation of subgrid scale fluxes, applied to both heat and moisture transport, allows to correctly reproduce moist convective boundary layers with mixed-phased clouds even with coarse grid resolution. Implementation of state-of-the-art microphysics scheme for GPU systems not only led to significant speedup of the computations, but in general improved the multi-GPU scaling of the model.

Variational Data Assimilation in the Constructor of Dynamic Soil Carbon Models

Wed, 21 Jan 2026 00:00:00 +0500

This work presents an automatic adjoint-model construction within the Carbon Cycle Model Constructor (CCMC) that enables variational data assimilation (VDA) for estimating the initial state of soil dynamic carbon models. The adjoint is generated once from the generic pool-flux representation used in CCMC, which allows efficient gradient evaluation and iterative optimization of the initial pool vector without constructing a model-specific adjoint. The proposed approach is tested with two soil carbon models: SOCS (Soil Organic Carbon Saturation) and RothC (Rothamsted model). Data assimilation experiments are performed using long-term field observations of soil carbon content. The entire VDA workflow, including the adjoint solver and optimization algorithm, is implemented in the same Fortran code base as CCMC. CCMC+VDA implementation is fully compatible with the MPI+OpenMP TerM land surface model and provides a reusable, scalable foundation for variational soil-carbon data assimilation on modern supercomputers.

Novel Oxygen-Deficient Centers and Other Intrinsic Defects in Amorphous SiO2: Quantum Molecular Dynamics Simulations

Wed, 21 Jan 2026 00:00:00 +0500

The formation of stoichiometric and oxygen-deficient amorphous states of a-SiO₂from a corresponding crystal was simulated using the melting-quenching procedure. To simulate the entire process, quantum molecular dynamics implemented in the VASP program and supercells containing 64 Si atoms and 128 O atoms were used. This size allows modeling the formation of rings with a small number of Si–O–Si bridges. At given heating and cooling rates of 0.5 K/fs, the transformation of the crystal’s atomic network into a disordered structure was studied depending on the melt stabilization temperature. It is shown that despite the strong change in the topology of the atomic network in a-SiO₂compared to the crystal, the bulk of the atoms constitute a continuous network of SiO₄tetrahedra linked by oxygen vertices. Local disturbances of this arrangement of atoms are intrinsic point defects of SiO₂, which are given special attention. It has been shown that most of the defects present in oxygen-deficient states are also present in stoichiometric states of a-SiO₂. Along with the known intrinsic defects of SiO₂, new defects were also identified, including those associated with oxygen deficiency. It is shown that for two generally accepted models of oxygen-deficient centers (ODCs) in a-SiO₂, the energy of formation of an oxygen vacancy is significantly lower than the energy of formation of a twofold coordinated silicon atom. The point defects of a-SiO₂identified in this work create a foundation for the interpretation of experimental data on the radiation resistance of optical fibers, the effects of high-power laser radiation on optical coatings, and in microelectronics, where insulating layers based on a-SiO₂are used.

Characterization of the Role of Amino Acid Residues in the 2-Hydroxybiphenyl 3-Monooxygenase Catalysis Based on Bioinformatic Analysis of the Flavin-dependent Monooxygenases and Supercomputer Modeling of the Structure of Mobile Fragments Applying Variational Autoencoders

Kirill E. Kopylov, Maxim A. Shchepetov, Vytas K. Švedas — Wed, 21 Jan 2026 00:00:00 +0500

By modeling of predominant conformations of mobile loops in previously unresolved regions of 2-hydroxybiphenyl 3-monooxygenase structure (PDB ID: 5BRT) using GPU-accelerated metadynamics simulations integrated with artificial intelligence and high-performance computing the full-length protein model was built. Combined with bioinformatic analysis of the flavin-dependent monooxygenases it allowed to propose the functional role of amino acid residues in the 2-hydroxybiphenyl 3-monooxygenase catalysis. Three subfamily-specific residues Glu359, Lys339, Arg360 and the Asp332 residue, conservative throughout the entire family of flavin-dependent monooxygenases, form salt bridges Glu359-Lys339 and Arg360-Asp332, which stabilize alpha helices preserving the integrity of the Rossmann fold of the FAD-binding domain; subfamily-specific residues Trp338 and Glu359 provide the correct positioning of alpha-helices by interacting with two conservative residues Asp557 and Arg555 from the hydroxylase domain.

NAD binding pocket is formed by a number of subfamily-specific residues Trp38, Ser40, Ser42, Arg46, Ser47, Ala180, Asn205, Ser291, Trp293 located in an elongated pocket adjacent to the FAD binding site. The Asp313 residue, conservative in the entire family of flavin-dependent monooxygenases, directly interacts with FAD through hydrogen bonding with 2’-OH-ribitol, contributing to the binding and orientation of the cofactor. The Arg46, Ser47, Gly202, Ser203, Asn205, Arg242, Val253, Trp293, Met321, and Pro320, conservative for the entire family, play a crucial role forming the substrate binding site. The binding of cofactors and substrate in a quaternary complex and their orientation due to interactions with subfamily-specific positions Arg46, Ala180, His181 and Trp293 allows to perform the hydride transfer to the substrate stereospecifically. The triple stacking interaction between the FAD isoalloxazine ring, NADH nicotinamide ring and the subfamily-specific residue Trp293 leads to the formation of a highly stable charge-transfer complex and preferential Pro-S position in 2-hydroxybiphenyl 3-monooxygenase catalysis.