Reducing Design Bottlenecks

Traditional surrogate-based design optimization (SBDO) is hitting a wall, especially with high-dimensional, complex designs. In this new paper, Dr. Namwoo Kang presents a next-gen framework using generative AI, integrating three key models: - Generative model (design synthesis) - Predictive model (performance estimation) - Optimization model (iterative or generative) Rather than optimizing directly in a high-dimensional design space (x), the workflow introduces a low-dimensional latent space (z) learned via generative models. ➡️ z → x → y z = latent variables x = CAD geometry y = performance (drag, stress, etc.) This means we’re no longer hand-coding design parameters or doing trial-and-error with simplified surrogate models. 🧠 Why this matters: - Parametric modeling is no longer a bottleneck - Complex shapes are learned directly from CAD - Dynamic and multimodal performance data (1D, 2D, 3D) can be used - Near real-time optimization is possible #AI #GenerativeDesign #CAE #DesignOptimization

Here's something nobody told me until I was four years into my freelance career as a hardware engineer: The parts library is the real bottleneck in PCB design. Not the circuit design. Not the layout. It's the parts. I've lost entire workdays hunting down the "correct" version of a connector in a library. I've watched boards get held up for weeks because of wrong manufacturer part numbers, EOL parts, or zero stock. I've seen quotes blow up mid-process because a chip came back NRND. These aren't rare exceptions. Given the number of parts in any BOM, they happen with high probability. So I wrote a breakdown of how to actually fix this. Six steps. Practical. No fluff. Link in comments.

🧠 What advantages does Generative AI (Generative Design) have over traditional design optimization? In my last post, I compared deep learning–based surrogate models with traditional ML approaches. This time, from a design optimization perspective, let’s look at how Generative AI (built on deep learning) differs from—and often surpasses—conventional design workflows. Below are five standout advantages. 🔹1. Non-Parametric Design Optimization Traditional optimization starts by parameterizing a 3D shape into a small set of 1D variables, which inevitably restricts the design space. With Generative AI, we learn a latent representation z from the raw 3D geometry x, then search for the optimal z and decode it back to the design x. This enables non-parametric shape generation unconstrained by hand-crafted parameters. The generator can be coupled directly with CAE to evaluate objectives and constraints; when sufficient CAE data exists, a predictive model can be trained and paired with the generator for even faster loops. 🔹2. Real-time Inverse Design Conventional predictive models learn the forward map x → y (design → performance). What designers really want is the inverse: given a target y, find the optimal design x. Generative AI enables this via conditional generation, producing optimal designs that meet target specs in (near) real time. 🔹3. Synthetic Data Generation Building accurate predictive models requires many (x, y) pairs, yet industry often lacks diverse design examples x. Generative AI can create synthetic 3D designs, which can be automatically evaluated with CAE to produce labels y. This jump-starts dataset creation and alleviates the most common bottleneck in training predictive models. 🔹4. Realistic Rendering for Instant Customer Feedback A practical benefit of Generative AI is rapid, photorealistic rendering of optimized concepts from simple text prompts. You can show customers lifelike visuals immediately, collect feedback, and steer the design toward what users actually prefer—bridging aesthetics and performance early. 🔹5. Seamless Design–Engineering Integration Generative AI closes the loop between design and engineering. From a rough concept sketch, it can produce 3D geometry and renderings and estimate engineering performance via predictive models. Instead of waiting for long feedback cycles, designers get live engineering guidance and can iterate toward optimal solutions themselves. We’ve summarized these ideas in the attached visual PDF. 👇 #DesignOptimization #GenerativeAI #GenerativeDesign #DeepLearning #EngineeringDesign

Speed is the only variable. Engineering is the rate at which we turn a hypothesis into a physical reality. That rate is currently throttled by the computational tax of traditional simulation. We have built incredible tools for CFD (Computational Fluid Dynamics) and FEA (Finite Element Analysis). They are the bedrock of modern engineering. But they are slow. A high-fidelity simulation can take hours or days to converge. Reasoning from first principles: 1. The Bottleneck: Design quality is a direct function of iteration cycles. 2. The Physics: Traditional solvers scale with complexity. As designs get more complex, the "wait time" grows exponentially. 3. The Solution: Large Physics Models (LPMs). By training neural networks on the fundamental laws of the universe—Navier-Stokes and Maxwell’s equations—we move from calculation to inference. The Data: 🎯 Speedup: LPMs offer a 10,000x to 100,000x increase in simulation speed compared to traditional numerical solvers. 🎯Accuracy: We are seeing error rates as low as <1% compared to high-fidelity ground truth. 🎯 Efficiency: We can now explore 100,000 design permutations in the time it once took to run one. This isn't about replacing the engineer; it’s about increasing the bandwidth of human ingenuity. When the "compute tax" drops to near-zero, the engineer is no longer a technician waiting for a progress bar. They become a pure architect of intent. We reach frontier technologies—fusion energy, zero-emission transport, and advanced materials—years ahead of schedule. AI is the turbocharger for the human mind. Let's build. #AI #PhysicsML #Engineering #DeepTech #FrontierTech #Innovation #FirstPrinciples

Engineering Velocity: Reflections on Designing and Building Automotive Body Dies with Minimum Time and Cost After decades in tool engineering, I’ve learned that reducing die lead time comes from eliminating unpredictability across the classic workflow Design, Simulation, Machining, Assembly, and Tryout. When these stages act as a continuous process rather than isolated steps, both time and cost fall naturally. In design, stabilized geometry, controlled radii, and simplified addendum build the foundation for predictable forming. Excessive beads and over-correction might seem safe, but they usually turn into machining hours and extended tryout loops. In simulation, accuracy depends on disciplined inputs material curves, friction, binder pressure. A closed-loop cycle, where compensation updates flow directly into CAD and NC programming, prevents fragmentation and brings the die closer to its real forming behavior before steel is cut. During machining, multi-stage strategies and CAD-driven toolpaths tighten accuracy and cut rework. When the compensated model drives NC directly, machining becomes execution rather than interpretation. In assembly, modular interfaces standardized shoes, pillars, and pockets—reduce adjustment time and make the die’s mechanical behavior more predictable in spotting. Finally, tryout confirms the truth of every upstream decision. Press dynamics and material variability still require refinement, but when the digital preparation is coherent, tryout becomes calibration rather than rescue. Real reductions in time and cost come not from shortcuts, but from continuity when design, simulation, machining, assembly, and tryout reinforce one another with technical discipline and practical insight.

Developers are no longer the bottleneck. Are designers becoming one? I do things to prevent that. The rate of AI advancements in dev work is more noticeable than in design. The asymmetry is real. Cursor, Claude Code, Codex... the tools are compressing implementation time fast. A dev team that used to need 3 weeks to implement a design now needs 3 days. Design hasn't kept up because it's much harder to automate. At least at the quality bar that matters. Code either works or it doesn't. Design • has to work, • AND be right for the user, • AND fit the brand, • AND hold up in edge cases. If the handoff is still a week, the review cycles remain the same, the process will not change, we're going to have a problem. Most design teams haven't noticed it yet. When shipping was slow, we had a buffer. There was always time to iterate, refine, get feedback. Now that buffer is gone. The velocity gap is real, and it's widening. This already has some effects: • Companies are already running leaner design teams. • PMs are producing wireframes with AI. • Devs are shipping UIs without a designer in the loop. • Design is getting absorbed by whoever is moving fastest. Designers might become a bottleneck. Here's what I'm doing to avoid becoming one: 1. I learn to ship, not just design. I picked up Lovable as a weapon, and I taught myself to go from concept to working prototype without a dev. 2. I've moved upstream into decisions. Thinking about what to build and why is much more important when execution is easy. 3. I audit my work regularly. I ask myself, if a PM with AI can do 80% of what I did last month, what's the 20% that requires my expertise? That's my real value. That's it. Are you a designer who's already felt this shift? What changed first? ✌️

Every design org I walk into has the same problem. They just describe it differently. "Our designers are drowning — 4 people covering 10 scrum teams." "We need a NorthStar experience but the team is stuck in feature factory mode." "I think we need to hire a CDO but I'm not sure what I'd even ask them to do." Different words. Same issue: the design function has outgrown the way it's organized, and nobody has the time or vantage point to figure out what needs to change. That's the work I do now. I've led design orgs from 5 to 285 people — at ServiceNow, GE Digital, Compass, Cisco's security business, and Google Cloud. I know what a healthy design org looks like because I've built them. I come in as a fractional CDO and within two weeks assess the whole picture — talent, process, toolchain, strategy. Not with a framework someone sold me at a conference. With an approach I call workflow archaeology: understand how work actually moves through the organization before you change anything. The real bottleneck is almost never what leadership thinks it is. At one company, everyone assumed they needed more designers. The actual problem was that every decision required a round trip through Figma, a handoff spec, a dev build, and a review cycle. We collapsed that by moving to code-first design. Capacity improved because we removed a broken process, not because we added headcount. At another, the team was talented but directionless — no shared vocabulary between product, design, and engineering. My partner Jorge Arango and I ran a two-day ontology workshop before touching a single tool or process. Everything after moved faster because people were finally solving the same problem. Here's what makes this moment different: the boundaries between design, engineering, and product are dissolving. AI can generate production-ready components. PMs can prototype without waiting for a designer. Engineers can explore UX alternatives in code before a spec exists. Most organizations haven't rethought who does what, or why. That's an organizational design question, not a tooling question. Most design orgs don't need a transformation. They need someone who will look at how work actually happens and make targeted changes that unlock the team — a toolchain shift, a staffing model change, rethinking how disciplines collaborate, or helping leadership understand what design should actually be doing for the business. The companies that get this right don't hire a full-time executive and wait six months for a strategy deck. They bring in someone who's seen the patterns and can start making things better in weeks. If your design org is stuck and you're not sure why, I'm happy to talk. Greg Petroff is a fractional Chief Design Officer with experience at Google Cloud, ServiceNow, GE Digital, Compass, and Cisco. He partners with Jorge Arango through Unfinishe_ (unfinishe.com).

How to Find the Real Bottleneck in Your Review Process Do not start by changing review rules. Start by looking at where review time actually goes. For one week, read the comments. Look at the back and forth. Notice what reviewers spend time on. That tells you what to fix first. If reviews focus on style, formatting, or missing basics, automate it. Add linting, formatting, tests, and simple policy checks to CI. Humans should not be catching things a machine can catch faster. If reviewers keep saying the change is too big or hard to follow, the problem is batch size. Require smaller pull requests. Use feature flags so unfinished work can still merge safely. Small changes move faster and break less. If reviews turn into design debates, alignment is happening too late. Pull requests are not the place to decide approach. Add a short prework step for risky changes that explains the problem, the plan, the risks, and how to undo it. If reviewers hesitate because they are worried about production, you have a release problem. Add gradual rollouts, fast rollback, and clear ownership during deploys. When shipping feels safe, reviews speed up. Do not try to fix code review directly. Fix what code review is compensating for.

If you own cross-functional delivery, this is the cleanest way to trade risk for time.     Here is the why. When you can model thermal, fluid, and control behavior in one place, you stop waiting on scarce components and long test cycles. You explore design choices in hours, not quarters. Quality rises because edge conditions are no longer hidden. Cost drops because the first physical build is already close to right.     This is not theory. A cold-chain manufacturer used advanced system simulation to extend shelf life for temperature‑sensitive products, improve vaccine refrigeration performance, and cut development and test time by up to 80 percent. They modeled refrigerant absorption in compressor oil to secure temperature autonomy, swapped in parts with shorter lead times using supplier parameters, and shared easy-to-use models so mechanical designers could run studies without specialists. Fast system-level results under a week earned leadership support that months of CFD and prototypes could not.     The pattern I recommend is simple: start with a system model that any engineer can run. Include the few parameters that reflect your real bottlenecks: heat load, compressor behavior, and component lead times. Use it to answer one business question each week, then publish the decision and inputs so the team can reuse it on the next program.     If you adopt one change this quarter, make simulation models accessible to non-experts and tie them to supply constraints. That’s how you turn uncertainty into plan‑able work.     If this would help your team, pick one product area and pilot a simple system model for next week’s gate review. 

Business constraints are designs new bottleneck. In my last post, I got some positive feedback on the patterns I’m seeing in our work and with customers. Judgement has emerged as an important skill in design leadership. → Taste is sensing what feels right or off → Skill is what allows you to make it → Craft is using skill and taste under constraints → Judgment is deciding what to do next But I'm seeing something bigger. Yesterday, I was reviewing AI-generated concepts for a commenting system. They were working…seemed fast and functional. But it didn’t feel like we could move forward on any of the work because we hadn’t defined the constraints tightly enough. We had a functioning AI concept in under an hour. The team can easily get excited, but as we evaluated the screens, it became clear that the constraints weren’t tight enough, and the business intent was getting lost in the experience. That’s when it clicked for me that the new bottleneck is understanding how to set tighter business constraints on the problem in this fluid, iterative design process.  I’m seeing this with our clients as well. Budgets are still budgets. Legal is still legal. Decision making is still happening in the same way. We talk a lot about how design is disrupted by AI, but I’d argue that biggest stressor that design and product teams face now is figuring out how to align the challenges with stakeholders. Craft produces outputs. Judgment interprets outcomes. Constraints decide what survives. Strategy is choosing what not to do. Design impact isn’t just aesthetic, and it isn’t even just strategic. It’s bounded by capital, timing, politics, platform limits, and risk tolerance. Most stakeholders really aren’t arguing about pixels. Most of the time, they focus on reducing risk or on alignment. Now here’s the interesting part in an AI world: AI accelerates craft. AI influences judgment. But constraints are not accelerating at the same rate. Before: Craft → weeks Judgment → quarterly reviews Constraints → annual planning Now: Craft → hours Judgment → same-day Slack threads Constraints → rolling reprioritization Without judgment, AI floods the system with options. And without constraints, teams get stuck chasing noise at an ever-quicker pace. That compression increases volatility. Even my team is feeling it ZURB A lot of designers think craft is the bottleneck to business success. I don’t think that’s true. I understand why it feels that way. It’s uncomfortable to figure out. If craft isn’t the constraint, then design can’t hide behind polish or execution anymore. It has to step into a different role. It has to facilitate. It has to communicate. It has to align people around the business constraints. Craft isn’t the ceiling. Judgment under constraints is what limits impact.