Adithya Ramachandran

Doctoral Researcher

• My collaboration partner during my time is Diehl Metering GmbH, Nuremberg, Germany
  
  

  Reflection: My doctoral research has been highly influential in shaping my perspective on AI solutions and their real-world impact. I particularly enjoyed working with smart meter data at the household level, navigating its intricacies and challenges. Analyzing thousands of time series data simultaneously, each with unique characteristics, influencing parameters, and psychographic implications helped me refine my analytical approaches. I found the instances when developing AI solution such as high forecasting errors coinciding with school closures and reopenings, unique consumer demand behaviors through unsupervised clustering to be obvious in hindsight and intellectually rewarding. Working in tandem with other data modalities and integrating disparate data silos to uncover valuable insights also helped be develop parallel thinking. While working with GIS data in urban logistics networks, I often spent long hours scouring maps to understand complex networks and their metadata, especially when tackling problems where key entities remain hidden beneath the surface. Developing analytical solutions for such issues and devising validation schemes was both challenging and motivating. Beyond the hands-on work, I thoroughly enjoyed engaging with fellow researchers who enriched my knowledge and inspired new ideas that I was eager to explore. These experiences significantly broadened my domain knowledge, not only in AI but also in the respective domain. As an academic researcher, I had the freedom to explore problems independently while also learning when to seek guidance and support—an invaluable lesson in both research and personal growth.

• Digital Twin Development: Developed a Digital Twin for urban heat and water supply networks that synergizes multi-modal data (smart meters, SCADA, GIS) into a unified knowledge graph serving as a central knowledge base. Utilized Python for data processing, NetworkX for network analysis, and Neo4j for graph-based inference. Developed a web application with HTML, CSS, and JavaScript for interactive visualization, enhancing infrastructure monitoring and decision-making.
• Data Engineering & Data Pipeline: Engineered a scalable data pipeline to process and harmonize smart meter and GIS data. Managed ~17,500 meters with hourly readings spanning up to 5 years, alongside an SQL-based utility database over 30,000 infrastructure components, and other public resources. Applied analytical and relational techniques to extract meaningful features, enabling ML-driven solutions for varying downstream tasks.
• Advanced Time Series Forecasting: Developed deep learning models for multi-horizon heat and water demand forecasting using advanced feature representations (time-frequency analysis, seasonal decomposition, multi-variate relationships), achieving up to 95% accuracy. Improved forecasting by 10%-40% across rural and urban districts compared to statistical and deep learning baselines. Leveraged attention mechanisms to reduce model size by 97% while maintaining accuracy, enabling optimal resource management, cost savings, and reduced emissions.
• Unsupervised Representation Learning: Constructed a contrastive representation learning framework to identify distinct consumption patterns across residential, commercial, and agricultural sectors using smart meter data. Analyzed variations in consumption of individual consumers and across different consumers, enabling behavioral insights, region-specific demand analysis, and proactive grid optimization.
• Anomaly Detection & Infrastructure Optimization: Designed analytical and ML-driven anomaly detection models for real-time monitoring such as leakage detection and localization, aiding predictive maintenance and optimizing grid efficiency.
• Technical Stack: Python, SQL, GIS processing, Graph Neural Networks, Wavelet Transforms, Contrastive Learning, JS/HTML/CSS.

Intern - Thesis

Reflection: My time at BASF marked my first experience in a large-scale industrial setting and provided an opportunity to bridge academic knowledge with real-world applications. Transitioning from C++ academic projects to industrial implementation was insightful. Additionally, I gained a deeper appreciation for the importance of confidentiality, particularly in handling internal tools and proprietary technologies. This period also coincided with the onset of the COVID-19 pandemic. Just two weeks into my master's thesis, lockdown measures were implemented, requiring me to quickly adapt my plans, working environment, and discipline.

• Architected a high-performance C++ simulation framework integrating PDEs for finite element algorithms, to model shrinkage in Fibre Reinforced Plastics (FRPs) to enable fast, accurate simulations to assist consumers in formulating the optimal FRP designs for their needs..
• Implemented and analysed numerical methods for multi-physics modeling, and stress-strain relationships for composite material behavior analysis.
• Technical Stack: C++, FEM, deal.ii, Abaqus, Moldflow, Fiber Reinforced Plastics, Non-linear material behavior.

Research Associate

Reflection: Working at CNDE, IITM was my first exposure to academic research and its ecosystem. Coming from a mechanical engineering background, it allowed me to take a multidisciplinary approach by integrating programming into mechanical applications. Writing physical equations governing material behavior and solving them numerically was an exciting challenge. This experience brought many first-time realizations, such as the speed-up in execution through vectorization, understanding marching in time and space with respect to numerical methods, using symbolic math to precompute values instead of performing matrix multiplications, and running overnight simulations leading up to my first conference. This period not only deepened my appreciation for programming but also gave me the confidence to pursue a master's degree in computational engineering.

• Developed a finite element modeling package using MATLAB and C, implementing vectorized computations and sparse matrix solvers to simulate acoustic wave propagation in polycrystalline materials through custom discretization schemes.
• Formulated and implemented mathematical solutions for PDEs, including finite difference time-domain (FDTD) solvers with stability conditions, demonstrating understanding in numerical methods, signal processing, and highperformance computing to support in-house academic research.
• Technical Stack: MATLAB, Finite Element Methods (FEM), Finite Difference Method (FDM), Non-Destructive Testing (NDT), Ultrasonic wave propagation, Polycrystalline materials, Non-linear material behavior, Simulation, Abaqus, COMSOL.

Master of Science in Computational Engineering

Student Assistant

Institute of Applied Mechanics, Friedrich Alexander University, Erlangen, Germany (July 2018 – September 2019)

• Worked on a MATLAB/C++ based code with deal.II library for Multiscale Simulations of Polymers.
• Implemented the Capriccio method for coupling of Continuum Mechanics and Molecular Dynamics for further FE analysis.

Projects

• Solar Panel Defect Classification
  • Designed a Deep Learning framework based on Resnet using PyTorch for classification of defects in solar panels (3rd place in University Competition with respect to F1 score).
  • Incorporated Pretrained models of ResNet-18, ResNet-32, and DenseNet.
• Advanced Programming
  • Hybrid CPU-/GPU-Parallelisation (High End Simulations in Practice) Developed a hybrid CPU-GPU simulation method for Molecular Dynamics using C++ and CUDA through domain partitioning and dynamic load balancing.
  • Forward Simulation and Optimization of StarCraft 2 Worked on a Forward Simulator to verify and simulate a given StarCraft 2 build list using C++ and also create an optimal build list using genetic algorithm for various gameplay scenarios.
• Others
  • K-Nearest Neighbor, Logistic and Linear Regression, Support Vector Machine (SVM) and Gaussian Mixture Models (GMM), K-Means, Mean-Shift Clustering, Random Forests and Hidden Markov Models (HMM), Fully Connected, Convolution Neural Network (CNN), Recurrent Neural Networks (RNN).