Noticias e investigación.
Estrenos y proyectos, y las notas técnicas de nuestro grupo de investigación (publicadas en inglés). Ver la serie de vídeos →
Forschungszulage Dokumentation: The Defensible R&D Record
Forschungszulage Dokumentation explained: what the BSFZ and the Finanzamt each read, timesheets, contract research invoices and German-only annexes.
IEC 61508 and Machine Learning: The Part 3 Route
IEC 61508 and machine learning: what Edition 2 and the Edition 3 drafts say, why proven in use fails, and how Part 3 covers runtime, tools and weights.
Edge AI Functional Safety: The Silicon Half of the Case
Edge AI functional safety beyond the model: soft errors in NPUs, what safety islands certify, and where the safety monitor has to run on the chip.
Kalman Filter and Neural Network Hybrid State Estimation
Kalman filter and neural network hybrid state estimation: four ways a network corrects the filter, what each keeps of the guarantees, measured gains.
Functional Safety and Machine Learning: Who Holds the SIL
Functional safety and machine learning under IEC 61508 and TR 5469: why the learned controller never holds the SIL, and what evidence it still owes.
Research Allowance Germany: Contract Research in 2026
Germany's research allowance on contract research in 2026: 70% of the fee, a 20% overhead flat rate on top, 35% for SMEs, and three dates that decide.
Physics-Informed Neural Networks for Bioprocess States
Physics informed neural networks for state estimation in bioprocesses: what a PINN can infer, how to test estimability, and the measured results.
Physics-Informed Neural Networks in Model Predictive Control
Physics informed neural networks for model predictive control: the 74 percent compute saving, where PINN surrogates break, and how to build one.
Moving Horizon Estimation vs Kalman Filter: Which to Use
Moving horizon estimation vs Kalman filter: where the EKF fails on constrained nonlinear processes, what MHE costs in compute, and how to choose.
What Is Quantization Aware Training? QAT Explained
What quantization aware training is, how fake quantization and the straight-through estimator work, and when post training quantization is already enough.
What Is Edge AI Deployment? Definition and How It Works
Edge AI deployment explained: running inference on the device, how it works, when it is worth it, and what changes when the model acts on a live process.
Neural Differential Equations in Python: Pick a Solver
Neural differential equations in Python: torchdiffeq, diffrax, torchdyn and torchode compared, and which solver survives a stiff physical system.
Industrial Soft Sensors: What Happens After Deployment
An industrial soft sensor is accurate on the day it ships and less accurate every month after. What the measured decay looks like, how to tell a bad model from a bad day, and what it costs to keep one alive.
Neural Ordinary Differential Equations Explained
Neural ordinary differential equations explained through a physical process: the derivative as the model, the solver as the depth, and what it costs you.
Disadvantages of Reinforcement Learning on Real Plants
The disadvantages of reinforcement learning that only show up on a live process: data cost, sim-to-real failure, reward hacking, constraints and drift.
Safe Reinforcement Learning via Shielding, Explained
Safe reinforcement learning via shielding: how a shield is built, what it assumes, where it breaks, and when a barrier-function filter fits better.
Reinforcement Learning vs PID Control: What the Data Shows
Reinforcement learning vs PID control, measured on real rigs: where RL wins, where PID still holds, and why most plants end up running both.
Contract Research R&D Credit: 65%, 70% or 100%?
How the R&D credit treats contract research in the US (65%), Germany (70%), France and Spain, who claims it, and why the contractor's location decides.
Contract Research and Development Organization Types
A contract research and development organization is not always a pharma CRO. Compare CROs, CDMOs, RTOs, universities and independent R&D firms.
Waste Gasification Control Strategy: From an Operator to a Learned Model
An operator, a PID loop on air, a soft sensor for tar, MPC or a model that learns the feedstock: how to choose a waste gasification control strategy, with the measured numbers behind each rung.
Edge AI Deployment Strategies: Cloud, Gateway or Machine
Edge AI deployment strategies for models that act on a live process: what runs at the machine, the gateway or the cloud, and how to decide.
Digital Twin for Process Control: How Good Is Good Enough?
A digital twin for process control can check code, tune loops, predict inside MPC or gate setpoints. What fidelity each job needs before you trust it.
Bioreactor Control System: The Full Stack, Layer by Layer
What a bioreactor control system is made of: sensors, regulatory loops, supervisory control, data plane and learned control, plus a spec checklist.
AI for Industrial Process Control: What Must Be True First
AI for industrial process control comes in four authority levels, from prediction to moving an actuator. What each needs before a plant should trust it.
What Is Contract Research? Three Meanings, One Test
What is contract research? The pharma CRO, the university research contract and the independent R&D contractor compared: scope, IP and publication.
Difference Between R&D and Engineering: Where the Line Falls
The difference between R&D and engineering is uncertainty, not job title. Where the UK, OECD and German BSFZ draw the line, and how to document it.
Edge AI Deployment Pipeline: Model to Certified Controller
An edge AI deployment pipeline for models that act on a live process: budget, runtime, conversion, measurement, packaging, validation and rollback.
Post Training Quantization vs Quantization Aware Training
Post training quantization vs quantization aware training: what each changes, what the measured gaps look like on edge hardware, and when QAT pays.
Bioreactor Control Strategy: Open Loop to Learned Models
Open loop, cascade, MPC or a model that learns the growth curve: how to choose a bioreactor control strategy, with measured numbers from real deployments.
Zero Shot Sim to Real Transfer: When to Skip Calibration
Zero shot sim to real transfer means deploying a simulation-trained controller with no real data. Four conditions decide when that claim is defensible.
Neural Differential Equations for System Identification
Neural ordinary differential equations for nonlinear system identification: far lower open loop error than state space models, paid for in inference time.
Domain Randomization for Sim to Real on a Slow Process
How domain randomization for sim to real transfer works outside robotics: which factors to randomise on a slow process, and where the ranges come from.
AI in Chemical Process Control: Four Layers, Four Bars
AI in chemical process control sits at four layers, from soft sensing to the valve. Each has its own evidence bar. The results, sorted by layer.
Reinforcement Learning vs MPC for Process Control
Reinforcement learning vs MPC for process control: the published results only look contradictory until you sort them by regime. Which one is yours.
Control Barrier Function vs Model Predictive Control
Control barrier function vs model predictive control: what each guarantees, the failure each owns, what each costs per step, and when to layer them.
Control Barrier Function Safety Filter: How To Install One
A control barrier function safety filter wraps a policy you already have. Which filter family to install, the one tuning parameter, and the time budget.
Control Barrier Function Explained: The Safety Filter
Control barrier functions explained: what h(x) and the safe set mean, why the CBF quadratic program filters a controller in real time, and what it costs.
Sim to Real Transfer on a Live Process, Not a Robot
Sim to real transfer outside robotics: what the reality gap means on a slow, partly observed plant, and why randomisation alone does not close it.
Safe RL With Control Barrier Functions: What Breaks
Safe reinforcement learning with control barrier functions: what the guarantee actually promises, and the four assumptions that fail on a real plant.
Safe Reinforcement Learning for Industrial Process Control
Safe reinforcement learning for industrial process control: the five things that have to be proven before a learned policy is allowed to move a valve.
Neural Controlled Differential Equations for Irregular Data
Neural controlled differential equations for irregular time series, judged as a control problem: the mechanism, the benchmarks, and what breaks online.
German R&D Tax Credit for Contract Research: 70% Rule
The German R&D tax credit covers 70% of a contract research fee, only the client may claim it, and the contractor must be seated in the EEA. How it works.
Model Compression for Edge AI: Meet the Latency Budget
Edge AI model compression judged on measured wall clock, not FLOPs: the ordered prune, quantize, distil pipeline, and what a control deadline changes.
Contract Research and Development Services: A Buyer Guide
Contract research and development services outside pharma: what you receive, how a programme is priced under uncertainty, and how to evaluate a partner.
What Is Contract Research and Development? How It Works
Contract research and development explained: the definition beyond pharma, how a programme is scoped, who owns the foreground IP, and who claims relief.
Knowledge Distillation vs Quantization: How to Choose
Quantization changes the numbers, distillation changes the model. What the measured edge comparisons say, and how to pick under a real latency budget.
