Eugeen Katkovsky - Academia.edu (original) (raw)

Uploads

Papers by Eugeen Katkovsky

Wikipedia [2] defines it as a Technological Singularity (TS). Technological singularity is a hypo... more Wikipedia [2] defines it as a Technological Singularity (TS). Technological singularity is a hypothetical moment in the future when technological development becomes fundamentally uncontrollable and irreversible, which generates radical changes in the nature of human civilization. Similar or similar definitions exist in other literature. At the same time, no explanation is given about the specifics of the various technologies and their controllability. There is no ranking of the list of technologies in which the expectation of a singularity is the most desirable and quite easy to implement. Even more vague are the arguments about some artificial intelligence (AI) created by man, which will later improve itself and surpass the human mind in its ability to think and create new, even more perfect, minds. The prospect of the inhumane development of AI begins to frighten society with its uncontrollability and creates real public tension and rejection of AI. This leads to the conclusion that it is necessary to find a tool that should manage the process of achieving local TS in each technology. Another conclusion can be drawn about the sequence (i.e., non-simultaneity) of achieving local TS in all technologies.

The author in his developments on Artificial Neural Networks (ANN) uses a special software system... more The author in his developments on Artificial Neural Networks (ANN) uses a special software system NeuroSolutions which has a developed graphical user interface (NeuroSolutions GUI). (What architectures and structures of neural networks can be built using NeuroSolutions package can be found at http://www.neurosolutions.com/neurosolutions). NeuroSolutions satisfies virtually all requirements for automating the design, training, and delivery of ANN results in a practically applicable form. The NeuroSolutions GUI is based on a "Breadboard". ANN simulations are built and managed on mockups. In NeuroSolutions, designing a neural network is very similar to prototyping an electronic circuit. In an electronic circuit, components such as resistors, capacitors, and inductance coils are first lined up on the layout. NeuroSolutions instead uses neural components such as Axons, Synapses, and Attachments. The components are linked together to form a circuit. The circuit transmits electrical current between its components. The circuit (i.e., neural network) of NeuroSolutions transmits activity between its components, and is called a data flow machine. When the circuit is validated, data is entered and the system's response is examined at various points. An electronic circuit would use an instrument, such as an oscilloscope, for this task. The NeuroSolutions network uses one or more of its components within the research family (e.g., MegaScope). ANNs are built on a layout by selecting components from palettes, imprinting them on the layout, and then linking them to form a network topology. Once the topology is established and its components formed, the simulation can be controlled. An example of a functional layout is shown below: Representing neural networks as graphs has one drawback: the graph will not show how the network works. For example, variational autoencoders (VAE) look exactly like a simple autoencoder (AE), while

Neural network control of acceptance criteria of VVER-1000 in case of thermal hydraulic parameters uncertainties , 2018

The paper describes a methodology to perform a fast evaluation of the influence of some thermal-h... more The paper describes a methodology to perform a fast evaluation of the influence of some thermal-hydraulic input uncertainties, such as hydraulic resistance coefficients of the reactor core and of the reactor pressure vessel (RPV), onto the important thermal hydraulic distributions such as coolant temperature, pressure, DNB, void fraction, maximal fuel temperature, maximum cladding temperature etc. All these safety related parameters determine the operational state of the active core of the VVER-type reactors and are a part of the core acceptance criteria list that must be checked. The successful application of the Artificial Neural Networks (ANN) methodology to the above mentioned purposes is outlined. The ANN is trained applying a set of test calculations performed with the system code ATHLET in the frame of the OECD/NEA transient Kalinin-3 Benchmark. Analysis of the efficiency of different ANN architectures is performed based on the accuracy and on the convergence properties of the particular choice of ANN. For the learning procedure of ANN a random variation of a double sided 95% confidence interval of different hydraulic resistance coefficients in the RPV is used. For each set of the hydraulic resistance coefficients a test simulation is performed with the system code ATHLET to obtain the corresponding thermal hydraulic distributions (mass flow, pressure, temperature) within the RPV. It is shown that for a particular class of stationary RPV thermal hydraulic parameter distributions ANN results are in a good correspondence with results obtained from the ATHLET simulation for such fluid objects of the active core, which have not been considered in the training sequence of the artificial neural network.

A detailed ATHLET [2] model of the inner structure of reactor pressure vessel is modeled using a ... more A detailed ATHLET [2] model of the inner structure of reactor pressure vessel is modeled using a special preprocessor which is able to generate automatically a three dimensional nodalization of the facility [1]. To tackle the problem of a large number of thermohydraulic variables and a large simulation CPU time, several enhancements are introduced in a new version of ATHLET: 1) an algorithm of the ATHLET Jacobian preconditioner is optimized; 2) a new direct sparse matrix solver based on the KLU algorithm of Tim Davis is implemented which gives a substantial acceleration of calculations for a large number of control volumes and makes feasible transient simulations with a detailed (ten thousands of control volumes) description of the investigated facility. In particular, in case of approximately 60000 control volumes and connections (the model used to investigate the OECD/NEA standard problem[3]), the acceleration up to 750 times in the Jacobian preconditioner and up to 25 times in the simulation time on a single CPU-core is achieved [5-6] compared to the previous version of ATHLET. 3) Further acceleration is achieved using Unsymmetric-pattern Multifrontal Numerical Factorization algorithm (UMF) of the same author Tim Davis . It gives an additional speed-up of about 5 times in comparison with . Both methods are tested for the UPTF facility TRAM C1 experiment . For this system the number of control volumes and connections is approximately 30000. UMF algorithm gives an acceleration of 200% with respect to the KLU algorithm. Currently the parallel version of the algorithm for the GPU architecture is being developed.

A diagnostics of initiating events, states of reactor systems and relevant safety variables is im... more A diagnostics of initiating events, states of reactor systems and relevant safety variables is important for safe operation of an NPP. On the one side one needs to correctly interpret signals from measuring instrumentation. On the other side a fast classification of the transients significantly improves the reliability of NPP operation.

The paper presents a methodology which will enable in the future performing of fast transient ana... more The paper presents a methodology which will enable in the future performing of fast transient analysis of NPP with VVER-1000 reactors. It is based on artificial neural networks (ANN) method. For the training of the network the best-estimate system code ATHLET (GRS) is used with detailed nodalization (multi-channel model) of the Reactor Pressure Vessel (RPV). The present work is dedicated to the training procedure which is connected at this stage of development only with the setting in the system the RPV hydraulic resistances. Training samples are created by random variation from a double sided 95% confidence interval of 50 different hydraulic resistance coefficients in the RPV. For each set of hydraulic resistance coefficients a test calculation is performed with ATHLET to obtain the corresponding thermohydraulic distributions (mass flow, pressure, temperature) within the RPV. For this training procedure the total and the assembly power distributions used in all calculations remain unvaried and equal the nominal values, e.g. at this stage a dynamic 3D neutron-physics model is not being taken into account. As a result of the neural network training an application program is created. This program is capable to give almost simultaneous answer about all thermo-hydraulic fields at each point of the RPV. A comparison of the system results with results obtained from the AHTLET simulation gives a good correspondence for such nodes (fluid objects) of the active core, which have not been taken into account in the training sequence of the artificial neural network. The paper contains also some ideas how to apply the artificial network modeling of the RPV in the current practice.

Drafts by Eugeen Katkovsky

On the issue of compliance to requirements of IAEA of the RELAP-5 and Athlet codes, 2021

The new document of IAEA SSG-2 (Rev. 1) [8] postulates: provides recommendations and guidance on ... more The new document of IAEA SSG-2 (Rev. 1) [8] postulates: provides recommendations and guidance on the use of deterministic safety analysis (DSA) and its application to nuclear power plants in compliance with the requirements established in IAEA Safety Standards Series Nos SSR-2/1 (Rev. 1) [1] and GSR Part 4 (Rev. 1) [2]: "Deterministic safety analyses for normal operation, anticipated operational occurrences, design basis accidents and design extension conditions, including severe accidents, as defined in SSR-2/1 (Rev. 1) [1] and in the IAEA Safety Glossary [13], are essential instruments for confirming the adequacy of safety provisions.
In this work compliances of RELAP-5 MOD3 codes (certification passport of Rostekhnadzor (Russian Federal Service for Environmental, Technological and Nuclear Supervision) No. 180 of 28.10.04) and Athlet -2.1 (certification passport of Rostekhnadzor No. 350 of 17.04.14) to requirements of document SSG-2 (Rev 1). [8] are considered.
It should be noted that later versions of the RELAP-5 and Athlet codes have no differences in methodical approaches from above the specified versions and do not change conclusions of this article. Besides, all mentioned below belongs both to versions of the RELAP-5 3D codes and to Athlet CD.

Dear Colleagues! Please familiarize yourself with my doubts about the use of Relap-5 codes. I wou... more Dear Colleagues! Please familiarize yourself with my doubts about the use of Relap-5 codes.
I would be very grateful if I receive criticism of my judgments from you.

Neural network control of acceptance criteria of VVER-1000 in case of thermal hydraulic parameters uncertainties , 2018

On the issue of compliance to requirements of IAEA of the RELAP-5 and Athlet codes, 2021