Utafiti huu umechunguza kama miundo mikubwa ya lugha inaweza kutumiwa si tu kama zana za uboreshaji zinazozalisha wagombea wa nambari, bali pia kama wabunifu wa kidijitali wanaotafsiri taarifa za kihandisi zilizopangwa zinazotokana na uchanganuzi wa elementi finiti. Tatizo lililochunguzwa ni kupunguza uzito wa kisanduku cha bawa cha kompoziti ya kaboni/epoksi cha ndege ya kikanda ya abiria 90, huku uzito, uharibifu wa tabaka la kwanza, kielezo cha kuvunjika cha Hoffman, mikazo ya stringer na mipaka ya buckling vikizingatiwa kwa pamoja.

Mtafiti alitengeneza Semantic Design Report ambayo, baada ya kila uchanganuzi wa Nastran, haikuripoti tu objective function na thamani za constraints, bali pia ilieleza ni maeneo gani ya kimuundo yaliyo muhimu, ni design variables zipi zinazodhibiti maeneo hayo, athari ya variable moja kwa uzito, mahali pa modes za chini zaidi za buckling na maeneo yenye mzigo mdogo yanayoweza kupunguzwa unene. Miundo ya Claude Sonnet na Claude Opus ilisoma ripoti hii na kupendekeza muundo unaofuata wa wing box. Katika njia ya mfuatano, mgombea mmoja alitathminiwa kwa kila iteration; katika njia ya population, wagombea kumi walitathminiwa.

Matokeo bora zaidi ya LLM yalipatikana kwa njia ya mfuatano iliyotumia Claude Opus: kwa evaluations 201 za finite-element, uzito wa reference wing box ulipunguzwa kwa asilimia 28,42. Genetic algorithm ilipunguza uzito kwa asilimia 31,23 kwa evaluations 36.000. Hivyo LLM ilikaribia matokeo ya normalized mass ya genetic algorithm kwa tofauti ya percentage points 2,81 huku ikitumia evaluations za finite-element mara 179 chache zaidi. Hata hivyo, uwiano huu haumaanishi faida ya mara 179 katika real wall-clock time, matumizi ya nishati au gharama ya fedha; unalinganisha tu idadi ya finite-element solves.

Utafiti ulifanywa kwa geometry moja ya ndege ya kikanda, mpangilio mmoja wa tabaka za kompoziti uliowekwa na maneuver moja ya symmetric positive-g. Negative-g, gust, roll, ground loads, flutter, manufacturing defects na fuel–structure feedback ya kiwango cha ndege nzima hazikujumuishwa katika optimization loop. Utafiti pia ni preprint ambayo haijapitia peer review.

Swali kuu la utafiti ni lipi?

Classical structural optimization algorithms kwa kawaida hupokea kutoka finite-element analysis ya design moja objective value moja na constraint values chache. Kwa mfano, algorithm inaweza kuona mass ya wing box, highest damage ratio na lowest buckling factor; lakini haiwezi kuelewa moja kwa moja damage imetokea kwenye panel gani, design variable gani inadhibiti panel hiyo au ni eneo gani lenye margin kubwa ambalo linaweza kupunguziwa material kwa usalama.

Swali kuu la utafiti ni hili: ikiwa large language model inapokea baada ya kila solve detailed evaluation iliyobadilishwa kuwa lugha inayotumiwa na structural engineers, je, inaweza kufikia practical wing-box design kwa analyses chache sana kuliko classical black-box optimization?

Mtafiti alichunguza swali hili kupitia subproblems tatu:

• Finite-element results zinawezaje kubadilishwa kuwa report ambayo large language model inaweza kuelewa bila kupoteza engineering detail?
• Je, large language model inaweza kupendekeza candidates wapya kwa kutumia relationships kati ya design variable na stress, ply damage, buckling na mass?
• Approach hii inatumia finite-element solves chache kwa kiasi gani kuliko genetic algorithm yenye evaluations 36.000 inayoboresha wing box hiyo hiyo?

Kwa nini objective function pekee haitoshi?

Kwa genetic algorithm, taarifa kwamba “normalized mass ni 0,75 na worst constraint ni 0,98” inaweza kutosha kupanga candidates. Kwa large language model, namba hizi mbili hazielezi kwa nini design imefanikiwa au imefeli. Ikiwa model inasonga mbele kwa kubadilisha namba tu, inakuwa karibu zaidi na mutation operator inayofanya kazi kwa lugha ya kawaida kuliko designer anayeelewa physical system.

Approach iliyotengenezwa katika utafiti inaunganisha numerical result na structural cause. Kwa mfano, report inaweza kueleza kwamba low buckling factor inatokana na panel fulani kwenye outer wing skin, variable gani inadhibiti thickness ya panel hiyo na kuongeza variable hiyo kwa step moja kunavyoongeza normalized mass. Kwa hiyo model inaweza kuamua “ongeza nguvu kwa step moja kwenye mid-span region ambako buckling mode imejikusanya, na ondoa material kutoka slack tip region” badala ya kusema tu “ongeza unene kwa ujumla”.

Semantic Design Report ni nini?

Semantic Design Report ni structured engineering briefing inayozalishwa kiotomatiki baada ya kila static aeroelastic analysis na buckling analysis. Katika main methods section, report imefafanuliwa kwa blocks sita:

Katika contributions na results sections, report pia inaelezwa kuwa na region-based narratives. Kulingana na maelezo katika Appendix B, longer human-readable archive inaweza kuwa na full region table, component-level results na region narratives. Shorter version inayotumiwa ndani ya prompt imepunguzwa ili kuwasilisha critical information bila kuzidi context window.

Finite-element model iliundwaje kiotomatiki?

In-house framework inayoitwa Furor inabadilisha design vector yenye variables 19 kuwa Nastran model inayoweza kuendeshwa bila human intervention. Workflow ina stages nne kuu:

1. CAD geometry: OpenCASCADE ilitumika kuunda outer surface ya wing box, front na rear spars, ribs, spar caps na stringers.
2. Topological cuts katika generalized coordinates: Internal structures ziliwekwa kwa coordinate η katika span direction na ξ katika chord direction.
3. Meshing kwa GMSH: Regular quadrilateral elements zilitumika katika skin na spar webs; quad-dominant unstructured elements zilitumika katika rib na transition regions.
4. Nastran input file: Composite shells zilibadilishwa kuwa PCOMP, stringers kuwa CBAR, lumped masses kuwa CONM2 na load-distributing connections kuwa RBE3 cards.

Figure 1 inaonyesha rib, spar-cap na stringer lines katika wing bays mbili kwa coordinates η–ξ. Figure 2 inaonyesha structural categories sita ndani ya finite-element mesh; Figure 3 inaonyesha aeroelastic panel mesh na lumped masses zilizoongezwa juu ya structure hiyo hiyo. Kwa sababu ya persistent region labels hizi, hata design ikibadilika results zinaweza kuhusishwa na engineering regions zile zile kama “root spar web”, “mid-span skin” au “tip stringer”.

Aeroelastic analysis chain ilifanyaje kazi?

Kwa kila candidate, computational blocks tatu kuu ziliendeshwa:

Aerodynamic loads zilihesabiwa upya juu ya deformed wing kwa kila structural candidate mpya. Kwa upande mwingine, engine, fuel, fuselage, tail, payload, landing gear na system masses zilichukuliwa kutoka single multidisciplinary analysis iliyofanywa kabla ya optimization na zikawekwa fixed.

Lumped fuselage model iliyotumiwa katika comparison iliwakilisha fuselage kwa single rigid point mass kwenye center of mass badala ya detailed finite-element shells. Wing, horizontal tail na vertical tail zilikuwepo katika aeroelastic model, lakini internal structure ya wing box pekee ndiyo ilibadilishwa na optimization variables.

Sifa za ndege na wing box zilizochunguzwa ni zipi?

Wing box ni two-spar, two-span-bay structure. Kink point iko kwenye engine-pylon attachment. Skin, spar webs, rib webs na spar caps zimemodelliwa kwa composite shells; stringers kwa bar elements.

Composite layup structure ikoje?

Fixed quasi-isotropic layup imetumika katika skin, spar webs na rib webs:

\[ [0^\circ/+45^\circ/-45^\circ/90^\circ]_s \]

Subscript s inaonyesha kwamba layup ni symmetric kuhusu mid-plane. Katika Nastran model, orientation blocks nne zimefafanuliwa na kupanuliwa kuwa full eight-block laminate kupitia symmetry option. Nominal thickness ya kila prepreg ply ni 0,05 mm.

Spar caps na stringers zimetazamwa kama unidirectional carbon laminates zilizopangiliwa katika span direction. Utafiti haukufanya ply angles kuwa optimization variables; lengo lilikuwa kuelekeza comparison kwenye search behavior ya large language model na genetic algorithm.

Strength values zimetumika pamoja na asilimia 50 damage-tolerance reduction na fatigue knockdown factor. Katika skin na spar webs, minimum thickness ya 0,20 mm kwa kila orientation block imewekwa. Kwa symmetric eight-block structure, limit hii inalingana na total manufacturing floor ya 1,6 mm.

Ni design variables zipi zilitumika?

Optimization vector ina high-level variables 19:

\[ \mathbf{x}=(\mathbf{x}^{skin},\mathbf{x}^{spar},\mathbf{x}^{cap-w},\mathbf{x}^{cap-t},\mathbf{x}^{str},\mathbf{x}^{top}) \]

Makundi matano ya kwanza yana values tatu kila moja katika root, wing kink na tip control stations, hivyo kuunda sizing variables 15. Kundi la mwisho lina topology variables nne zinazofafanua rib count na rib distribution katika bays mbili.

Stringer counts katika inner na outer bay zimewekwa fixed kuwa tisa na nne. Ply orientations pia ni fixed. Kwa hiyo utafiti haujachunguza design variables zote zinazowezekana kwa composite wing box; umechunguza subproblem yenye variables 19 iliyo restricted lakini yenye maana ya kihandisi.

Rib positions ziliamuliwaje?

Rib positions katika kila wing bay zimehesabiwa kwa combination ya uniform distribution na root-clustered cosine distribution:

\[ \eta_k=(1-c_b)\frac{k}{N_b-1}+c_b\left[1-\cos\left(\frac{\pi k}{2(N_b-1)}\right)\right] \]

Hapa η_k ni dimensionless rib position ndani ya bay, N_b ni rib count, k ni rib index na c_b ni cosine-distribution weight kati ya 0 na 1.

• Wakati c_b = 0, ribs zinawekwa kwa spacing sawa.
• Wakati c_b = 1, ribs zinajikusanya zaidi kuelekea bay root ambako bending na shear loads ni kubwa.

Sizing values katika root, kink na tip stations zimeinterpolatewa linearly kwenye span, na kila panel region imepewa value ya midpoint yake. Composite thicknesses kisha zimezungushwa hadi manufacturable ply increments za 0,05 mm. Figure 6 inaonyesha continuous control-point curve inavyobadilishwa kuwa stepped ply-thickness distribution katika finite-element model.

Optimization objective ilifafanuliwaje?

Lengo ni kupunguza wing-box mass ikilinganishwa na manually sized initial design:

\[ \bar{m}_{WB}(\mathbf{x})=\frac{m_{WB}(\mathbf{x})}{m_{WB}^{(0)}} \]

Hapa m_WB(x) ni mass ya candidate wing box kwa kg, m_WB⁽⁰⁾ = 2.456 kg ni initial mass na m̄_WB ni dimensionless normalized mass. Kwa initial design, m̄_WB = 1.

Optimization problem imetolewa katika study kwa compact form ifuatayo:

\[ \min_{\mathbf{x}\in X}\bar{m}_{WB}(\mathbf{x}) \]

\[ g_i(\mathbf{x})\leq 1,\quad i=1,\ldots,n_g \]

\[ x_j^{Root}\geq x_j^{Kink}\geq x_j^{Tip},\quad j\in\{skin,\;spar\ web\} \]

Inequality ya mwisho ni structural taper rule inayohakikisha skin na spar-web thickness haziongezeki kutoka root kwenda tip.

Ni damage na safety constraints zipi zilitumika?

Element na mode constraints nyingi zimekusanywa kwa laini kwa kutumia Kreisselmeier–Steinhauser function. Utafiti hautoi mathematical expansion ya kina ya KS function, lakini umetumia smoothing parameter ρ = 80 kwa damage constraints na ρ = 30 kwa buckling modes tano.

Ni flight load gani ilitumika?

Wing box ilipimwa kwa maneuver moja tu ya symmetric positive-g pull-up:

Negative-g, discrete gust, rolling pull-up, ground loads na flutter analyses hazikujumuishwa katika comparison hii. Kwa hiyo optimum iliyopatikana si aircraft wing iliyothibitishwa kwa full certification load envelope.

LLMopt loop ilifanyaje kazi?

Katika kila iteration, large language model ilisoma current best feasible design, designs zilizojaribiwa hivi karibuni na Semantic Design Report. Model ilitoa new design variables, rationale ya move na candidate type katika fixed JSON schema.

Variables zote ziliwakilishwa si kwa engineering units bali kwa position ya asilimia 0–100 kati ya lower na upper bounds. Decoder ilibadilisha percentage kuwa physical value, ika-round hadi permitted manufacturing step na ikatumia root ≥ kink ≥ tip rule kiotomatiki. Missing variables zilichukuliwa kutoka current champion design na invalid variable names zikaondolewa.

Convergence state ya model ilitolewa katika kila user message. Champion isipoboreshwa kwa muda mrefu, broader exploration iliombwa; candidates kadhaa mfululizo zikiwa infeasible, smaller recovery moves ziliombwa. Adaptation hii haikufanywa kwa training ya model weights, bali kwa engineering instructions ndani ya iteration prompt.

Tofauti kati ya sequential na population methods ni ipi?

Sequential method ilitumia analyses chache zaidi; population method ilitumia Nastran workers kumi kwa wakati mmoja kwenye high-performance computing cluster, hivyo wall-clock cost ya iteration moja ikakaribia muda wa slowest candidate. Kwa hiyo “fewer FEM evaluations” na “shorter actual runtime” si dhana moja.

Genetic-algorithm comparison ilifanywaje?

Reference method ilikuwa NSGA-II algorithm kutoka pymoo library. Ingawa problem ilikuwa single-objective, NSGA-II ilitumika kama reference driver inayotoa feasibility-first constraint handling na genetic operators.

Kwa genetic algorithm, study imeripoti campaign moja tu ya random seed. Vivyo hivyo, campaign moja tu imeripotiwa kwa kila combination ya LLM model na method. Kwa hiyo run-to-run distribution, confidence interval au statistical superiority havijabainishwa.

Convergence curves zilionyesha nini?

Logarithmic evaluation axis katika Figure 8 inaonyesha wazi tofauti ya sampling budget kati ya LLM na genetic algorithm. Campaigns zote nne za LLM zilipunguza normalized mass kutoka 1,00 hadi karibu 0,80 ndani ya evaluations takribani 30 za kwanza. Katika early stage, safety margins kwenye constraints ambazo bado hazikuwa active zilitumika haraka.

Ply-strain na buckling constraints zilifika boundary region ndani ya takribani evaluations 20–50 za LLM. Baada ya hapo, kupunguza unene katika region moja kulisababisha strain au buckling violation, hivyo progress ikapungua na results zikaunda plateau karibu 0,72.

Genetic algorithm ilikaribia constraint boundaries zile zile baada ya takribani evaluations 10.000. Hata hivyo, kwa search ndefu zaidi ilifikia mass takribani percentage points tatu chini ya LLM designs.

Best mass results ni zipi?

Approximate mass difference kati ya best LLM result na genetic algorithm ni 69 kg. Katika normalized scale iliyotumiwa na study, difference ni 0,0281, yaani percentage points 2,81.

Evaluation-count ratio ni:

\[ \frac{36\,000}{201}\approx179{,}1 \]

Kwa hiyo study inatoa conclusion ya “FEM evaluations mara 179 chache zaidi”. Statement hii inalinganisha analyses 36.000 za genetic algorithm na analyses 201, reference ikiwa imejumuishwa, za best LLM run.

Structural constraints zilitumika kwa kiwango gani kwenye optimums?

Katika optimums zote, fiber-direction ply strain ilifikia asilimia 97,6–98,6 ya limit ya 2.000 µε. Lowest buckling factors zilikuwa 1,562–1,607, karibu na limit ya 1,5. Metrics hizi mbili ndizo zilikuwa active constraints za design.

Hoffman index ilibaki 0,707–0,728 na kuwa na safety margin takribani asilimia 27–29 chini ya limit ya 1,0. Stringer tension na compression constraints pia hazikuwa limiting. Hivyo methods zote zilifika structural boundary inayofanana, lakini zilisambaza material tofauti kwenye span ya wing.

LLM na genetic algorithm zilitengeneza designs tofauti vipi?

Figure 10 inaonyesha genetic algorithm na LLM designs zikifika kwenye same constraint boundary kupitia structural strategies tofauti.

• Genetic algorithm ilisambaza skin na spar-web thickness kwa balance zaidi kwenye span.
• LLM designs zilijikusanyia material katika wing-kink station karibu na engine pylon.
• Katika LLM designs zote nne, inner-bay rib count ilipunguzwa kutoka nane hadi nne; genetic algorithm ilitumia ribs tano.
• Outer-bay rib count ilikuwa 10–15 katika LLM designs na 12 katika genetic algorithm.
• Katika LLM designs, baadhi ya skin, spar web, spar cap na stringer variables katika kink station zilikaribia upper bound.
• Genetic algorithm ilipunguza root skin hadi takribani asilimia 41 ya reference, huku LLM designs zikiiacha kati ya asilimia 62–72.

Tofauti hii inaonyesha kwamba LLM ilitumia region narratives kwenye report kutengeneza more regional design logic kama “imarisha critical mid-span, punguza material katika regions nyingine”. Hata hivyo, haijathibitishwa kwamba strategy hii itatoa result ile ile kwa wing geometries nyingine.

Strain distribution katika Figure 11 inaeleza nini?

Figure 11 inaonyesha initial design, genetic algorithm na LLM designs nne kutoka root hadi wing kink kwa deformation iliyoongezwa mara tano. Katika designs zote, highest fiber-direction strain ilitokea katika general region ile ile, kwenye spanwise spar web katika wing root.

Common critical region hii inaonyesha kwamba optimization methods tofauti hazikubadilisha physical load path. Ingawa methods zilichagua material distributions tofauti, basic bending na shear loads katika positive-g maneuver ziliendelea kufanya wing root kuwa sehemu muhimu.

Initial value iliyo kwenye figure ni 1.895 µε, wakati Table 6 ina 1.896 µε. Difference ya microstrain moja haibadilishi interpretation ya results, lakini source text haijatoa rounding explanation.

Computational cost ya kutumia LLM ni ipi?

Sababu ya kuwa na calls nyingi kuliko iterations ni kwamba model iliitwa tena candidate iliyokwisha evaluated ilipopendekezwa upya. Katika sequential Sonnet run, duplicate candidates 32 zilitengenezwa tena; katika Opus run, sita. Hakuna invalid JSON au format error iliyoonekana.

Utafiti unaripoti token counts lakini haujakokotoa API cost, energy consumption au carbon impact ya LLM calls. Monetary costs za Nastran license, HPC queue na LLM service hazijalinganishwa katika common scale.

Ni conclusions zipi zinazoungwa mkono na utafiti?

• Kuunganisha finite-element output na design variables pamoja na physical regions kuliwezesha large language model kupendekeza feasible designs kwa analyses chache.
• Best LLM run ilipunguza initial wing-box mass kwa asilimia 28,42 katika evaluations 201.
• Ilikaribia result ya asilimia 31,23 ya genetic algorithm kwa percentage points 2,81.
• Finite-element evaluation count ilipungua kutoka 36.000 hadi 201, na sampling difference ya mara 179 ikapatikana.
• Sequential LLM method ilitoa mass kidogo chini kuliko population method kwa evaluations chache zaidi katika problem hii.
• Result difference kati ya Claude Opus na Claude Sonnet ilikuwa ndogo kuliko difference kati ya sequential na population methods.
• LLM na genetic algorithm zilifika strain na buckling boundaries zile zile kwa material distributions tofauti.
• Structured engineering feedback iliunda optimization interface yenye taarifa nyingi zaidi kuliko scalar objective value pekee.

Ni conclusions zipi ambazo utafiti hauthibitishi?

• Utafiti hauthibitishi kwamba large language models ni mara 179 efficient zaidi kuliko genetic algorithms katika structural optimization problems zote.
• Thamani ya mara 179 ni kwa FEM evaluation count tu, si wall-clock time, electricity consumption, monetary cost au total CPU time.
• Haijathibitishwa experimentally kwamba LLM result inaweza kutengenezwa au kuruka kwa usalama katika real composite wing.
• Haijaonyeshwa kwamba large language model inaweza kuchukua nafasi ya structural engineer.
• Single-run results si ushahidi wa statistical superiority kati ya methods.
• Haijaonyeshwa kwamba NSGA-II inawakilisha genetic, evolutionary au surrogate-based optimization methods zote.
• Haijaonyeshwa kwamba safe design imepatikana chini ya negative-g, gust, flutter na full certification loads.
• Kwa kuwa ply angles, stacking sequence na stringer count hazikuoptimized, full composite design space haijachunguzwa.
• Haijaonyeshwa kwamba LLM yenyewe inasolve physics laws au inachukua nafasi ya Nastran analysis; physical evaluation bado inafanywa na finite-element solver.

Hii ina maana gani kwa Uturuki?

Ujumbe muhimu zaidi kwa Uturuki ni kwamba AI inaweza kutumiwa katika engineering si tu kwa text generation au code assistance, bali kama decision layer inayofanya kazi juu ya existing CAE infrastructure. Wakati finite-element solver inafanya physical evaluation, large language model inaweza kutafsiri critical regions, safety margins na effects za design variables na kuamua analysis inayofuata ifanywe wapi.

Approach hii inaweza kuchunguzwa nchini Uturuki katika composite aircraft structures, UAVs, satellite na launch systems, automotive lightweighting, wind-turbine blades na high-cost analysis problems nyingine. Hasa katika problems ambako finite-element solve moja huchukua masaa au inahitaji software licenses nyingi, kupunguza analysis count kunaweza kuwa muhimu.

Kwa local application, design rules zinapaswa kuendana na manufacturing limits za engineering organizations za Uturuki, retrieval system salama kwa internal historical designs inapaswa kuundwa, results zihakikishwe kwa multiple load cases na physical tests, na LLM suggestions zipitie expert approval.

Mbinu na Matokeo ya Utafiti

Technical summary ya research design

Optimization strategies tano zilizolinganishwa

Main quantitative findings

1. Genetic algorithm ilipunguza normalized mass kutoka 1,0000 hadi 0,6877.
2. Best LLM result ilikuwa 0,7158 kwa Claude Opus sequential method.
3. Normalized-mass difference kati ya best LLM na genetic algorithm ni 0,0281.
4. LLM iliokoa takribani 698 kg kutoka reference mass; genetic algorithm takribani 767 kg.
5. Approximate dimensional mass difference kati ya LLM na genetic algorithm ni 69 kg.
6. Kuna difference ya mara 179,1 kati ya evaluations 201 za LLM na evaluations 36.000 za genetic algorithm.
7. LLM results zote nne zilikusanyika katika range 0,7158–0,7260, band nyembamba kuliko percentage point moja.
8. Sequential methods mbili zilitoa lower mass kuliko population methods za models zile zile.
9. Active ply strain ilibaki 1.951–1.972 µε katika optimums zote.
10. Lowest buckling load factor ilibaki 1,562–1,607 na juu ya limit ya 1,5.

Technical message ya graphs na figures

• Figure 1: Inaonyesha parametric placement ya ribs, spar caps na stringers katika generalized η–ξ coordinates.
• Figure 2: Inaonyesha skin, spar cap, spar web, rib na stringer finite-element categories karibu na wing root.
• Figure 3: Inaonyesha lumped masses na doublet-lattice aerodynamic panel mesh zilizoongezwa kwenye structural mesh.
• Figure 4: Inalinganisha detailed fuselage na lumped fuselage model na inaonyesha kwamba lumped model imetumika katika study.
• Figure 5: Inaonyesha data flow ya Furor model generation, SOL 144, SOL 105 na objective-constraint aggregation.
• Figure 6: Inaonyesha conversion ya root, kink na tip control values kuwa manufacturable stepped ply thickness.
• Figure 7: Inaonyesha closed loop yenye LLM provider, prompt builder, decoder, Furor, decision maker na history archive.
• Figure 8: Inaonyesha LLM methods zikifika karibu 0,72 katika evaluations 201–1.001, huku genetic algorithm ikishuka hadi 0,688 katika evaluations 36.000.
• Figure 9: Inaonyesha ply strain na buckling kama active constraints, Hoffman na stringer stresses kama secondary constraints.
• Figure 10: Inalinganisha design philosophies ambapo LLM inajikusanyia material katika wing kink na genetic algorithm inatumia distributed sizing zaidi.
• Figure 11: Inaonyesha deformation na fiber-direction strain distributions za designs sita kwenye same color scale.

Methodological strengths

• LLM na genetic algorithm zilitumia same design space, same Nastran model na same constraints.
• Finite-element model iliundwa upya kiotomatiki kutoka mwanzo kwa kila candidate.
• Topological region labels zilihifadhiwa kati ya designs.
• LLM outputs zilifungwa kwa fixed JSON schema.
• Invalid designs au designs zinazovunja monotonicity rule ziliprojectiwa kwenye nearest permissible point.
• FEM evaluation count, LLM calls na token consumption ziliripotiwa separately.
• Final designs zililinganishwa si kwa mass pekee, bali pia strain, damage, buckling na displacement.
• Software version, operators, population na random seed za genetic algorithm zilielezwa.

Methodological limitations

• Kila optimization configuration iliendeshwa mara moja tu.
• Genetic algorithm ilitathminiwa kwa random seed moja.
• Temperature parameter haikudhibitiwa katika Claude models na provider default sampling ilitumika.
• LLM zilizotumika ni closed hosted models; model weights na service behavior zinaweza kubadilika kwa muda.
• Decoder ilirekebisha kiotomatiki baadhi ya invalid suggestions, hivyo raw LLM success na whole-system success hazijatenganishwa kikamilifu.
• LLM ilijaribiwa tu na rich semantic report; hakuna direct ablation run ambapo models zile zile zilitumia scalar objective na constraints pekee.
• Hakuna direct comparison na CMA-ES, differential evolution, Bayesian optimization au surrogate-assisted methods.
• FEM model haikuthibitishwa ndani ya study kwa physical wing au subcomponent test.
• Aircraft geometry moja na load case moja tu zilitumika.
• Ply angles, stacking order na stringer count ziliwekwa fixed.
• Absolute wall-clock time, electricity consumption na monetary cost comparison hazikutolewa.
• Hakuna public Furor au LLMopt code repository iliyoelezwa.

Documentation differences zilizobainishwa katika text

• Peak strain ya reference design imetolewa kama 1.895 µε katika Figure 11 na 1.896 µε katika Table 6.
• Semantic Design Report imefafanuliwa kwa blocks sita katika methods section, lakini imefupishwa kwa core blocks nne katika results section.
• Katika blocks nne za results, region narrative inahesabiwa, huku katika main six-block list slack regions, header na aggregated constraints zikiwa sections tofauti.
• Ingawa text inasema small refinement moves hupendelewa kwa improving designs, initial recommendation kwa new champion ni asilimia 70 exploration na asilimia 30 refinement; policy text inaeleza ratio hii kama flexible starting point.

Dokezo la Chanzo na Mbinu

Jina kamili la asili la utafiti: Large language models as efficient designers of composite wing-box structures via semantic feedback

Mwandishi: Luca Pustina.

Mpangilio wa waandishi: Utafiti una mwandishi mmoja.

Equal first author: Hakuna tamko la equal first authorship au equal contribution.

Corresponding author: Luca Pustina.

E-mail ya corresponding author: luca.pustina@uniroma1.it

Taasisi: Department of Mechanical and Aerospace Engineering, Sapienza University of Rome, Via Eudossiana 18, 00184 Rome, Italy.

DOI:10.2139/ssrn.6952641

Publication platform: SSRN.

Official SSRN record date: 16 Juni 2026.

Publication year: 2026.

Journal: Hakuna peer-reviewed journal au final journal publication iliyothibitishwa.

Original publisher: Hakuna final journal publisher. Utafiti umeshirikiwa kwenye SSRN preprint platform ya Elsevier.

Aina ya chanzo: Preprint research article yenye finite-element analyses, structural optimization na large-language-model-controlled computational experiments.

Peer-review status: Utafiti haujapitia peer review. Kurasa ya kwanza na zinazofuata zina warning “This preprint research paper has not been peer reviewed”.

Official source link:https://ssrn.com/abstract=6952641

Funding na project context: Utafiti umeripotiwa kuwa research activity ndani ya Italy’s PNRR CN4 MOST Sustainable Mobility research program, Spoke 1 “Air Mobility”, project CN_00000023 na Work Package 5.

Conflict of interest: Mwandishi ametangaza kwamba hakuna known financial interest au personal relationship inayoweza kuathiri study.

Generative AI statement: Mwandishi amesema alitumia Claude kuboresha readability na conciseness ya manuscript, kisha akapitia na kuhariri content na kuchukua responsibility. Matumizi haya yanahusu writing process na ni tofauti na LLM optimization campaigns zilizojaribiwa katika article.

Code na data access: Furor na LLMopt zimefafanuliwa kama internal frameworks. Hakuna public code repository au open data link yenye campaign data yote iliyotolewa. Example Semantic Design Report na prompt sections zimetolewa katika appendices.

Maelezo haya ya kisayansi ya Kiswahili yameandaliwa kwa kuchunguza text, equations, tables, finite-element visuals, optimization graphs na appendices za study iliyopakiwa. Scientific content inategemea tu methods na results zilizoripotiwa katika study. External sources zilitumika tu kwa bibliographic verification ya title, author identity, affiliation, DOI, SSRN record, platform nature na publication status.

Main limitations ni kutokuwepo kwa peer review, kutokuwepo kwa physical experiment, kutumia wing box moja na load case moja, kutotoa independent repeats kwa kila optimization method, LLM models kuwa closed services, kutokuwepo kwa broader algorithm comparisons na kwamba result ya mara 179 inategemea FEM evaluation count pekee.

Results zinaonyesha kwamba structured engineering feedback inayohusishwa na physical regions inaweza kugeuza large language models kuwa effective candidate generators katika finite-element optimization loop. Findings hazipaswi kutafsiriwa kwamba large language models zinaweza kufanya certified aircraft design bila human supervision au kuchukua nafasi ya finite-element solver.