The typical perception of IT in industry is: 1C reports, legacy software, and no freedom of action. If digitalization does exist, it's implemented only to meet KPIs. But if you truly delve into the tasks of industrial developers, it becomes clear: they work on complex, interesting cases that impact plant operations and simplify engineers' work.
At SIBUR, we understand firsthand the importance of digitalization, as we are the largest polymer and rubber manufacturer in Russia. The polyethylene, elastomers, and plastics we produce are used in a variety of industries, from construction and automotive to food and medicine. Our production scale is enormous: 17 facilities, 75,000 pieces of equipment, and hundreds of process parameters. Managing such a complex and large-scale production facility is impossible without digital transformation.
So, at our IT cluster—Digital SIBUR—we started thinking about how to demonstrate that our developers' products have an impact on a vast industrial complex. So we decidedtogether with HabrTell us about the technologies and solutions we create and how this helps move industry forward.
Digital Twin of a Pyrolysis Furnace and RTO: How Autopilot Works in Production
Polymers are everywhere, from disposable cups and plastic bags to rocket nose cones and airliner wings. The most popular and widespread of these are polyethylene and polypropylene. These materials are based on long chains of chemically bonded simple molecules—ethylene and propylene. These are light, flammable gases that are virtually nonexistent in nature. Industrially, they are produced from natural raw materials—ethane, propane, or heavier saturated hydrocarbons—through pyrolysis.
Pyrolysis is an energy-intensive, high-temperature process that occurs in specialized reactors—pyrolysis furnaces. Optimal process control, in terms of monomer production and energy consumption, is impossible without a precise, fast, and robust reactor design.
Previously, we at Digital SIBUR relied on foreign licensed software products to digitalize production. But in 2022, everything changed: vendors left the Russian market, and support and licenses became unavailable. The software continued to work at existing production facilities, but licenses would expire within 2-3 years. The question arose of how to continue operations when the traditional software was no longer available. And then it became clear: it was time to create our own solution that would calculate optimal operating modes and help manage production in real time.
This is how an entire software complex was created – a digital twin of the pyrolysis furnace.It works in conjunction with RTO (Real-Time Optimization) and the Advanced Process Control System (APC, as it's known internationally). Its main task is to optimize production processes in real time. Raw material composition and external conditions can change, so the system constantly monitors indicators and performs calculations to determine the most cost-effective operating mode. After completing the calculations, the digital twin signals the optimal APC mode.The system controls the technological process in real time: less raw materials and less energy mean higher product yield and higher company profits.
In developing our own solution, we didn't simply pursue import substitution: we created a high-tech product entirely using open source. The model is implemented in Python using the TensorFlow, NumPy, and SciPy libraries. For deployment and operation in production, we used Docker, Redis, Celery, and Flask (REST API). Our model features GPU/CPU support, modularity, customization, and an open API. Integration with nanoCAD is also supported for easy 3D modeling of the pyrolysis furnace configuration.
The most challenging aspect of the project was creating a mathematical model of the pyrolysis reactor and developing algorithmic solutions that ensure the required performance. Logically, the model divides the pyrolysis furnace into two entities: the pyrolysis furnace coil and the radiant chamber. The coil is the "heart" of the reactor, where the product flow moves at transonic speeds and undergoes complex chemical transformations. The coil model is based on the laws of conservation of mass, energy, momentum, and component composition. To avoid turning the calculations into an endless supercomputer task, we used a one-dimensional approximation: this reduces the mathematical description to a system of ordinary differential equations (ODEs). The problem, however, remains nontrivial—the mechanisms involve hundreds of components and tens of thousands of reactions. Here, we deliberately avoided using quasi-steady-state simplifications; the technological specifics of furnace operation require solving an ODE with boundary conditions.As a result, we were able to achieve calculation times within a few seconds – almost on par with the best solutions on the market.Additionally, we applied techniques to reduce the dimensionality of the pyrolysis mechanism, “cutting off” reactions and components that do not affect the final result.
The radiant chamber is the "body" of the reactor, with dozens of gas burners and walls heated to over 1000°C. It supplies heat to the coil. For this chamber, we used rigorous models based on the equations of mathematical physics. The solution to these equations is based on the finite volume method. We searched for a common steady state for the two models (the coil model and the radiant chamber model) using a simple iteration method with some modifications. Essentially, we achieved a solution that fits the complex thermodynamics of the pyrolysis furnace into an algorithm that calculates quickly, accurately, and reliably.
For our production, this solution is completely transparent: no black boxes, no licenses.The economic impact of implementing a digital twin is approximately 670 million rubles annually for existing production facilities. In the long term, we estimate that we will be able to save an additional 400 million rubles annually for new production facilities.
Predictive diagnostics: when breakdowns don't take you by surprise
Industrial workers' worst nightmares are sudden equipment shutdowns, when something breaks mid-shift, schedules are disrupted, employees are stressed, and the company loses money. But how can you catch a breakdown before it happens?
Previously, SIBUR used an imported predictive diagnostics system, but in 2022, like many others, it was forced out of Russia on a one-way ticket. New production facilities could no longer be connected, and support was nonexistent. A replacement had to be found urgently. Of course, there were already IT solutions on the market that could report problems out of the box. But they were too expensive and didn't cover all the needs of our production. We began searching for more affordable options and realized it would be more efficient to develop such a solution ourselves—after all, we are experts in this field and have our own team of data science specialists.
Together with colleagues from SIBUR's Monitoring and Diagnostics Center (MDC), we began developing our own software—a machine-learning-based system.This system was supposed to monitor the equipment's condition 24/7 and warn of problems before they became critical.
First, we selected sensors and wrote high-level software that collects information and monitors the equipment's condition: it tracks temperature, vibration, pressure, and other parameters, and sends a signal if anything deviates from the norm.
At this stage, we studied how a healthy unit should behave: with each installation of the system at a new production facility, we retrained the model. Production conditions vary, and old model settings have very limited applicability. We collected real data from a year of equipment operation, filtered it, and used this data to train the model to a normal state. Ultimately, we obtained a digital "snapshot" of healthy equipment at a specific production facility.
Once trained, this system automatically monitors the equipment's condition. If the system detects the first signs of a deviation from the norm, it generates an anomaly alert. The CMD diagnosticians analyze this alert and, together with production specialists, develop recommendations on how to avoid malfunctions. This approach helps prevent accidents, plan repairs, proactively procure spare parts, and avoid financial losses from downtime.
Our system includes rules for automatically evaluating monitoring results, which help quickly identify the most important events and generate recommendations for the CMD team.If the parameter deviation is within tolerances, no inspection is required. If it is outside these limits, we generate a task for equipment inspection. Expert rules are also included, which, based on the dynamics of changes in a large number of parameters, help determine the possible cause of deviations and offer recommendations for their elimination. The system automatically assesses the task's priority: analytical algorithms assign a level from 1 (high) to 5 (low), taking into account the magnitude and nature of the deviations. This way, we determine the urgency of the task.
We currently use statistical mathematical models such as the modernized Similarity-Based Modeling (SBM) algorithm. This approach is based on the idea that objects with similar attributes exhibit similar behavior. In the future, we plan to fully transition to neural network models based on autoencoders and have already begun implementing them in our software.
Under the hood, the program uses a high-tech stack: AI/ML and Digital Twin. Everything is built on open-source Python libraries (SciPy, TensorFlow).To properly monitor equipment, it's important to understand what defects can occur and how they manifest themselves. That is, what data—direct or indirect—can reveal this in real time.To do this, we use simplified Failure Mode and Effects Analysis (FMEA)—a method that helps link potential defects and equipment failures to parameters collected in real time. This allows us to identify risks early and take action to mitigate them.
The end result: a 10–15% reduction in losses on monitored equipment. But most importantly, fewer emergency situations and greater predictability of equipment operation directly improves reliability and team peace of mind.
Digital Lead Generation: How AI Helps Find New Clients
Problem: the classic approach no longer works. Previously, in industrial B2B, especially on the scale of SIBUR, customer flow was maintained through old-school magic: SEO, SMM, business trips, and "sit at a conference and exchange business cards." But at some point, all of this stopped producing results, and during the pandemic, some of the sales channels used became completely inaccessible. Leads were dwindling, marketers were panicking, and salespeople were dejected. That's when the idea struck: let's bring in AI and teach it to find loyal customers for us?
Digital lead generation was designed to solve several problems at once: expand the sales funnel, help enter new markets, and find customers faster when launching new products.AI was supposed to add flexibility to sales processes: help quickly reroute leads and adapt to changing market conditions, and automate the funnel process from the first lead to the sale. This is how the "Digital Lead Generation" product was born—a recommendation system that doesn't just search for potential clients, but selects those who are truly ready to work with us.
As a basis for lead generation, we used a so-called "data lake"—a collection of various databases and sources that aggregates all the necessary information, including customs statistics and industry databases, such as SPARK and its equivalents in other countries. Using industry databases, we can automatically categorize companies by production segment and prepare descriptions that serve as benchmarks for finding potential clients.
The algorithm works like this: first, we extract data from this "lake," then all the information undergoes preprocessing: the data is cleared of obviously unrealistic values, duplicates are removed, and the records themselves are enriched with additional features. Then, "smart" filtering is activated: at this stage, it's important to consider the priorities set by marketers.The challenge is not just collecting a huge number of leads, but selecting the ones that are truly suitable.Ultimately, we narrow the sample down from a million potential leads to a few hundred warm ones. This data is then automatically supplemented with a description and contact information and sent to the CRM. Finally, everything is ready for the sales manager: a client profile, contact information, and a brief company description. All that's left is to call and sign the contract.
When we first developed the prototype, there were some skeptics on the team: "Yeah, right, a neural network will find us clients, right?" But then, a couple of months later, they were already thanking us for the "warm" leads that turned into deals, and the sales team had more time for strategy instead of cold calls.
Previously, the lead generation process was lengthy: an employee would search for and process leads manually, creating a master list that would then be sent directly to sales. Now the process has become more streamlined and efficient: leads are categorized as cold or warm and prioritized, more sources have been added, and all data is stored and processed in a single system. Plus, we now have a dedicated call center that calls leads before sending them to the sales department.
Digital lead generation has helped improve the sales department's internal processes and customer interactions. The company now has a streamlined lead management system and a digitalized customer journey with transparent analytics and reporting. The time from a lead request to a qualified lead has been reduced from three months to one, and the cost per lead has decreased by 41%. External clients have also seen significant benefits: employees are responding more quickly to incoming inquiries, and small companies now have the opportunity to collaborate directly with SIBUR.We were able to reduce the time it took to collect and analyze data, filter out irrelevant leads, and that translates into hundreds of hours of saved work for sales and marketing specialists.Moreover, we gained leads in new areas and regions. The final figures we're proud of: over 28,000 leads and 1,500 new clients in a year and a half.
Where do ideas come from?
Digital lead generation is one of the areas where we use AI. Since launching our digital transformation, we've seen a growing number of hypotheses using neural networks within the company.To test them without financial risk, we opened an "AI Lab" at Digital SIBUR.New artificial intelligence ideas are submitted to the team. Each idea is evaluated to determine whether there is sufficient data, how much money can be saved or earned, and how relevant the solution is. The team then develops an MVP, tests it, and decides whether the idea has potential or should be shelved. Priority is given to solutions that will bring the greatest benefit to the company and are scalable. If the hypothesis passes all levels of testing, it is developed into a full-fledged project that is implemented into production.
When we launched the Lab, requests to test such product hypotheses began coming in from all business functions of the company, and now the AI Lab registry contains over 330 ideas.We have already implemented more than 30 cases, and about 140 hypotheses are in development at various stages.Below we will tell you about a project that was born from a hypothesis tested at the AI Lab.
Treasurer Co-Pilot: Managing Liquidity with AI
Let's imagine a huge industrial company, where every ruble can turn into two, or even disappear if the wrong decision is made. Now let's add some context: every morning, the finance department opens Excel with dozens of tabs to prepare forecasts: manually, with copy-pasting, using formulas like =IF(ERROR(...)). This approach is bad for both employees and the business: forecasts are time-consuming, and they're prone to errors that prevent effective investment to generate profit. Then the business asked for a user-friendly assistant that would generate more accurate forecasts, minimize errors, and free up financial analysts' time for more creative tasks (for example, figuring out where to invest the company's available cash).
This is how "Treasurer Co-pilot" was created—a software solution that helps make 90-day liquidity forecasts.In simple terms, the Treasurer's Co-pilot is an artificial intelligence system trained on historical data to forecast cash flows—receipts and expenses—with a certain degree of accuracy.
Our cash flow forecasting solution is based on a deep analysis of time series: we carefully sorted all the data by counterparty and expense and income item down to every trend, seasonal, abnormal, and other component. Only then did we assemble the forecasting system.Under the hood, we have a cascade of models: multi-season Prophet plus a set of linear regressions tailored to the floating payment schedule of counterparties.The data is complex: it includes weekly, monthly, quarterly, and annual seasonality (some additive, some multiplicative), as well as "conditional" seasonalities dependent on the calendar, such as the end of a month or quarter, tax dates, and holidays. The trend is broken, and the patterns are non-stationary, so a simple linear model isn't enough. For validation, we use a time series split with an expanding window: new data arrives every day, and this strategy has demonstrated the best metrics compared to others. Everything goes into production through MLFW—our internal service for securely running and monitoring ML models. There, we regulate the frequency of runs, monitor quality, and, if necessary, quickly deploy updates. We also used SAP ERP and BW for data collection, a data lake on Vertica, and FineBI for beautifully plotting forecast graphs.
The economic impact of implementing the Treasurer Co-pilot exceeded our expectations—365 million rubles in the first six months of this year.
Conclusion
We've finally completed our journey into the world of IT at SIBUR. We've shared some of our projects—they clearly demonstrate how digitalization simplifies operations at large production facilities at various levels, from lead generation to equipment health monitoring. Each product from our IT cluster represents a real-world problem solved using modern approaches, from RTO to ML, and these projects represent a complex, technically demanding, and highly engaging challenge for our team.
We believe that IT in industry is when your code makes the work of thousands of people easier and faster: preventing production accidents and helping your finance colleagues increase the company's free cash flow. This is where technology drives the industry forward—and that's true Heavy Digital. And if you'd like to ask about our technologies and projects, please don't hesitate to ask—we'll be happy to answer them!
