Exploring Collaboration and Communication Challenges in ML Engineering Teams: Insights From the Semiconductor Industry
Aidin Azamnouri, Markus Haug, Lucas Woltmann, Manuel Fritz, Justus Bogner, Stefan Wagner · 2026
Read Exploring Collaboration and Communication Challenges in ML Engineering Teams: Insights From the Semiconductor Industry by Aidin Azamnouri, Markus Haug, Lucas Woltmann, Manuel Fritz, Justus Bogner, Stefan Wagner (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-07389.txt.
Introduction
A. Azamnouri and five co-authors study teamwork problems inside machine learning teams in the semiconductor industry. Such teams mix engineers, data scientists, and domain experts with different goals and jargon, which makes long-term rollout harder. Hardware settings add strict data rules, long cycles, and tight coupling to plant processes.
Talking with twelve staff at a global chip company, the team found sixteen common challenges, with unclear roles and duties ranked most serious. They also collected habits staff found useful against those problems. The findings match known issues in cross-field work, but show how hardware constraints reshape them.
Catalog record
| CREATOR | Aidin Azamnouri, Markus Haug, Lucas Woltmann, Manuel Fritz, Justus Bogner, Stefan Wagner |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 104 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.07389
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Exploring Collaboration and Communication Challenges in ML Engineering Teams: Insights From the Semiconductor Industry Aidin Azamnouri, Markus Haug, Lucas Woltmann, Manuel Fritz, Justus Bogner, Stefan Wagner Abstract The integration of machine learning (ML) into complex software systems has increased challenges in collaboration and communication (CoCo) of the teams building these systems. ML engineering (MLE) teams often involve diverse roles, ML engineers, data scientists, software engineers, and domain experts, each bringing unique goals, experiences, and jargon. These interdisciplinary dynamics can make it challenging to deploy, reproduce, and maintain ML-enabled systems over the long term. Previous studies have uncovered several CoCo challenges and practices, but most have focused on software-centric companies, leaving limited empirical understanding of how these dynamics unfold in hardware-centric contexts. In hardware-centric environments, CoCo challenges are shaped by additional constraints such as strict data governance, long development cycles, and tight coupling with physical processes, which amplify coordination complexity and reduce flexibility. To strengthen empirical understanding in such settings, we present a qualitative investigation of MLE teams within a global semiconductor company, where ML-enabled systems and manufacturing processes introduce additional complexity. We interviewed 12 practitioners regarding CoCo practices, tools, challenges, and approaches. Through analysis, we identified 16 recurring challenges, with unclear roles and responsibilities emerging as the most critical, and common practices and recommendations practitioners considered effective in mitigating CoCo problems. While grounded in a single organizational context, our findings align with known issues in interdisciplinary ML-enabled systems development, but also demonstrate how these challenges manifest differently under hardware-driven constraints. Our results highlight directions for future research and tool support to strengthen CoCo in MLE projects and ensure the success of ML-enabled systems. [orcid=0009-0004-8409-2316] [orcid=0000-0001-9377-0677] [orcid=0000-0003-0720-8878] [orcid=0000-0003-4640-8477] [orcid=0000-0001-5788-0991] [orcid=0000-0002-5256-8429] keywords Collaboration ,Communication ,Machine Learning ,ML-Enabled Systems ,ML Engineering Team ,Semiconductor † † corresponding: Corresponding author 1 Introduction Bringing machine learning (ML) into real-world systems is not limited to designing novel algorithms, but involves a range of activities such as data collection, data preprocessing, model training, deployment, and continuous maintenance. These activities require close interaction between practitioners from different disciplines, including ML engineers, data scientists, software engineers, and domain experts. Hence, effective collaboration and communication (referred to as CoCo in this paper) become essential to align expectations, coordinate interdependent tasks, and ensure that ML-enabled systems can be successfully developed, deployed, and maintained over time ( 21 ; 28 ; 26 ) . Prior studies have explored challenges in software engineering for ML systems and machine learning operations (MLOps) in general ( 3 ; 25 ; 17 ; 16 ) , and highlight the complexities of integrating ML into software systems ( 41 ; 20 ) . This emphasizes the need for effective interdisciplinary collaboration among data scientists, ML engineers, and software developers ( 28 ; 26 ; 10 ; 6 ) . While ML engineering (MLE) in software-centric companies already requires the collaboration of many different roles ( 3 ; 10 ) , this becomes even more problematic in the semiconductor industry. Data scientists, domain experts, software engineers, sensor and imaging specialists, optics engineers, physicists, and hardware specialists, each with differing assumptions, vocabularies, and priorities, need to collaborate effectively. Misalignment in team understanding about goals, data ownership, or model deployment processes can introduce friction, increase technical debt ( 38 ; 32 ; 33 ) , and impede the effective translation of ML prototypes into production systems ( 44 ; 23 ; 12 ) . Such issues are compounded by the rapid pace of innovation and the critical demand for high reliability in semiconductor production pipelines ( 24 ; 43 ) . Recent research has underscored the growing importance of socio-technical factors in ML deployment ( 3 ; 37 ; 26 ) , highlighting how CoCo barriers, such as unclear team roles, mismatched expectations, and inadequate feedback loops, can hinder model adoption and maintenance. However, most of these studies focus on software companies. The semiconductor domain, characterized by proprietary data, stringent precision requirements, long development cycles, and very rigid deadlines, presents unique CoCo challenges that remain underexplored. Conducting a study within a global semiconductor company is therefore important for understanding CoCo challenges in high-tech manufacturing industries. Moreover, while previous work, such as 28 or 10 , explored CoCo challenges and practices of MLE teams in the software-centric industry, very few studies have deeply analyzed CoCo challenges across various MLE team members of a single hardware-centric company. Hence, our study aims to contribute a stronger empirical understanding of CoCo phenomena in contexts where ML integration intersects with high-stakes, precision-driven production processes. To start closing this gap, we studied the CoCo of the MLE teams of a global hardware-focused company via interviews. This company is a leader in semiconductor manufacturing technologies. It operates in a highly complex, multidisciplinary domain where teams of software engineers, data scientists, optical physicists, mechanical engineers, and systems integrators must work closely together. This environment exemplifies the intersection of advanced engineering disciplines with strict performance and quality demands, making effective collaboration essential yet difficult to achieve. In this company, there may be weeks or months of production time for a single product due to the very high requirements regarding the accuracy of the production. Training ML models in this context poses additional challenges in terms of sample size. For instance, the company generates large volumes of imaging data that are processed through a modular ML workflow designed for defect detection and classification in customer products. The organization adopts both data-driven and process-driven manufacturing approaches to continuously optimize production processes, thereby improving overall product quality and efficiency. Fulfilling this commitment demands highly customized solutions, which are developed in-house through close collaboration among data scientists, software engineers, and physicists. In the studied hardware-centric environment, ML projects are typically embedded within larger software–hardware systems rather than developed as standalone components. These projects follow a socio-technical structure in which ML solutions are tightly coupled with physical manufacturing processes, sensor systems, and production software. As a result, project organization, workflows, and system architectures differ substantially from those commonly reported in software-only ML settings. The workflow usually starts from a production or manufacturing problem, e.g., defect detection, process optimization, or quality inspection, rather than from a purely data-driven question. Domain experts define constraints and performance requirements grounded in physical processes and hardware limitations. Data collection and labeling are strongly influenced by sensor availability, measurement precision, and data governance constraints. ML model development is conducted iteratively, often with small or highly imbalanced datasets, requiring close feedback loops between ML practitioners and domain experts to validate assumptions and intermediate results. Integration into production systems involves software engineers embedding trained models into existing pipelines, ensuring compatibility with hardware interfaces, latency requirements, and reliability standards. Deployment is therefore not a single handover step but a negotiated process that spans development, validation, and gradual rollout. Architecturally, ML components are typically embedded as modular services within larger software systems that interface directly with hardware. A common pattern is a layered architecture consisting of data acquisition layers connected to sensors and machines, data preprocessing and feature extraction pipelines, ML model components for inference or decision support, and downstream software systems that consume ML outputs for visualization, control, or optimization. These components are connected through APIs, shared data repositories, or message-based interfaces, and must