From pencil to gen­er­a­tive AI: the evo­lu­tion of engi­neer­ing design

Tra­di­tional Drafting

For cen­turies, engi­neers began every concept with a sketch. A few lines on paper were enough to struc­ture ideas, estimate pro­por­tions, and visu­al­ize an outcome before anything was built. Before com­put­ers, designs were drawn by hand using tra­di­tional instru­ments such as rulers, com­passes, and pro­trac­tors. Yet freeform shapes, such as a ship’s bow, could not be captured pre­cisely with these rigid tools.

To draw smooth curves at full scale, engi­neers used flexible wooden strips known as splines, held in place by lead weights called ducks. Between these anchor points, the spline nat­u­rally formed the smoothest possible curve, achiev­ing both pre­ci­sion and elegance. The drawing board served not just as a work­space but as a physical exten­sion of an engineer’s thought process.

Computer-Aided Design (CAD) Modeling

As com­put­ers entered the design process, engi­neers began to study the physical prop­er­ties of splines so they could be modeled with math­e­mat­i­cal pre­ci­sion and repro­duced dig­i­tally whenever needed. The iter­a­tive def­i­n­i­tion of basis func­tions in NURBS (Non-Uniform Rational B‑Splines) aligned per­fectly with the increas­ing com­pu­ta­tional power of early computer systems, enabling levels of com­pu­ta­tional cal­cu­la­tion, pre­ci­sion, and B‑Rep (Boundary Rep­re­sen­ta­tion) modeling far beyond human capa­bil­ity. The tran­si­tion from main­frame com­put­ers to mini­com­put­ers and even­tu­ally to personal com­put­ers in the 1980s made computer-aided design (CAD) acces­si­ble to a much wider audience. The mouse replaced the pencil, becoming a new exten­sion of the human mind. It offered enhanced pre­ci­sion, digital stan­dards, and file formats that stream­lined work­flows and reduced manual errors.

While CAD rev­o­lu­tion­ized effi­ciency, it did not fun­da­men­tally change the under­ly­ing approach of design. Similar to tra­di­tional drafting, CAD still relied on the designer’s judgment to ensure overall con­sis­tency by placing geo­met­ric prim­i­tives or digital anno­ta­tions within a virtual drawing space. The concept of layers in CAD reflected the additive logic of the tra­di­tional drawing board, now trans­lated into the digital realm.

Para­met­ric Modeling and Algo­rith­mic-Aided Design (AAD)

In the 1990s, major advance­ments in 3D modeling and para­met­ric design intro­duced feature-based solid modeling. These devel­op­ments sig­nif­i­cantly changed how design­ers and engi­neers approached geometry, giving rise to modern para­met­ric and algo­rith­mic-aided design (AAD).

Algo­rithms began to generate geome­tries directly. When an inte­grated editor was used within CAD or other modeling software, objects were no longer manip­u­lated by hand or with a mouse. Instead, they were defined through pro­ce­dures expressed in pro­gram­ming languages.

This approach, known as script­ing, was rel­a­tively new to design practice. It rede­fined the con­nec­tion between the designer’s idea and the final output. The result was no longer a static digital drawing but an inter­ac­tive digital model that could respond dynam­i­cally to changes in input, manip­u­lat­ing the entire system accordingly.

The com­bi­na­tion of human judgment, computer pro­gram­ming, and engi­neer­ing methods expanded creative and ana­lyt­i­cal capa­bil­i­ties beyond what could be achieved manually.

Sim­u­la­tion-Driven Design (SDD) with CAESES

CAESES®, intro­duced in 2001, provides a spe­cial­ized CAD envi­ron­ment for para­met­ric modeling and inte­gra­tion with sim­u­la­tion tools. It includes its own script­ing language, the Feature Pro­gram­ming Language (FPL), which allows users to automate modeling tasks, create custom features, and link external tools directly to the design process.

Anything that can be executed in batch mode can be con­nected to CAESES, making it par­tic­u­larly effec­tive for coupling with CFD codes and other sim­u­la­tion software. It also offers a complete opti­miza­tion frame­work for explor­ing and refining the design space through Design of Exper­i­ments (DoE) and advanced opti­miza­tion algorithms.

This process defines sim­u­la­tion-driven design (SDD), where sim­u­la­tions guide engi­neers toward optimal solu­tions based on data and per­for­mance rather than intu­ition. For the past 25 years, sim­u­la­tion-driven design has led the way in advanc­ing modern engineering.

Future with Data-Driven Engi­neer­ing (DDE)

The future is increas­ingly shaped by the rise of gen­er­a­tive AI. Arti­fi­cial intel­li­gence has long been familiar to engi­neers through methods such as linear regres­sion, machine learning, and sur­ro­gate modeling, but in recent years, the rapid growth of gen­er­a­tive AI and large language models (LLMs) has shown that AI-based systems and intel­li­gent agents are becoming an essen­tial part of engineering.

The focus is grad­u­ally shifting from sim­u­la­tion-driven design (SDD) toward data-driven engi­neer­ing (DDE). Engi­neers today often possess exten­sive sim­u­la­tion data that requires sig­nif­i­cant com­pu­ta­tional time to produce. It is not effi­cient to rerun new sim­u­la­tions for every design vari­a­tion, espe­cially when these com­pu­ta­tions can take hours or even days and require access to high-per­for­mance com­put­ing systems or large clusters.

When data is already avail­able, the goal is to develop a machine learning-based sur­ro­gate model that can instantly retrieve results, even on a standard laptop. When data is limited, engi­neers can use Design of Exper­i­ments (DoE) algo­rithms, such as a Sobol sequence, to explore the design space effi­ciently, per­form­ing as few sim­u­la­tions as possible to build accurate sur­ro­gate models. These models then enable quick pre­dic­tions and design eval­u­a­tions without the need for repeated costly simulations.

This approach also enhances col­lab­o­ra­tion among col­leagues. Once the setup is complete, users do not need to be CFD experts to explore design alter­na­tives. With a well-prepared para­met­ric model, they can adjust para­me­ters and imme­di­ately observe how per­for­mance metrics change.

However, one must always consider the limits of extrap­o­la­tion. The accuracy of machine learning and sur­ro­gate models depends on the quality of the data, the coverage of the design space, and the nature of the under­ly­ing physical problem.

There­fore, a human expert remains essen­tial to evaluate and validate the outcomes, and across all stages of design and man­u­fac­tur­ing, the process remains human-centric. The engineer con­tin­ues to think, decide, and create, while the medium evolves to meet new chal­lenges. Each new gen­er­a­tion of tools, from pencil to code to gen­er­a­tive AI, expands the creative and ana­lyt­i­cal power of the human mind, enabling design with greater intel­li­gence, cre­ativ­ity, and understanding.