Towards intelligent methodologies for uncertainty quantification in civil nuclear energy safety

17 h In e na ional Con e ence on P obabilis ic Sa e y Assessmen and Managemen & Asian Symposium on Risk Assessmen and Managemen (PSAM17&ASRAM2024) 7-11 Oc obe , 2024, Sendai In e na ional Cen e , Sendai, Miyagi, Japan Towa ds in elligen me hodologies o unce ain y quan i ica ion in ci il nuclea ene gy sa e y Yu Chen, Edoa do Pa elli* Cen e o In elligen In as uc u e, Uni e si y o S a hclyde, Glasgow, UK Abs ac : Redundancies and physical sepa a ion o sa e y sys ems a e used in nuclea sys ems in o de o cope wi h po en ial componen ailu es, ex eme e en s and h ea s all cha ac e ised by la ge unce ain ies. Such unce ain ies a e una oidable as hey a ise om, o ins ance, manu ac u ing ole ances, modelling capabili ies, and spa si y in da a (e.g. one-o -a-kind sys em, ailu e da a), ex e nal and uncon ollable ac o s, e c. His o ically, unce ain ies in nuclea sec o ha e been ea ed in a highly conse a i e manne , wi h la ge, ine icien ma gins o ailu e. Quan i ying he e ec o he unce ain y is essen ial o ensu ing he sa e y o nuclea ins alla ions and also o suppo ing he li e ime economic iabili y o new nuclea powe plan in design, building, ope a ion and decommissioning. In ac , a p ope quan i ica ion and p opaga ion o unce ain y ac oss mul i-physical componen s allows o de e mine ulne able componen s, p io i ise in es men s, iden i y ope a ional ma gins and adop ele an measu es o gua an ee sa e y whils educing he o e all cos o ad anced nuclea design. Con en ionally unce ain y quan i ica ion was limi ed o semi- analy ical app oaches and equi ed s ong assump ions (e.g. Gaussiani y) due o he unmanageable compu a ional cos s o ull p obabilis ic assessmen s posing se ious ques ion on he alidi y o he esul s. Such adi ional me hods, which a e a om op imised, o en lack a igo ous p ocess o p opaga ion o unce ain ies, no mally esul ing in o e -enginee ing. Recen ad ances in in elligen compu ing b ings inspi a ion o new pe spec i es and analy ics in he way we design, build, ope a e and decommission ou sys ems. This pape p esen s an o e iew o he s a e-o - he-a me hodologies and ools o managing and quan i y unce ain y in nuclea sys ems. Keywo ds: Digi al Twin, Unce ain y quan i ica ion, Imp ecise P obabili y, Nuclea sa e y. 1. INTRODUCTION His o ically, he ea men o unce ain ies in nuclea analysis me hods has been ea ed in a highly conse a i e manne , wi h la ge, ine icien ma gins o ailu e. Quan i ying he e ec o he unce ain y is essen ial o ensu ing he sa e y o nuclea ins alla ions and also o suppo ing he li e ime economic iabili y o new nuclea powe plan in design, building, ope a ion and decommissioning. His o ically unce ain y quan i ica ion was limi ed o semi-analy ical app oaches and used o s ong assump ions (e.g. Gaussiani y) due o he unmanageable compu a ional cos s o a ull p obabilis ic assessmen . Such adi ional me hods, which a e a om op imised in he de elopmen o nuclea componen s, o en lack a igo ous p ocess o p opaga ion o unce ain ies, no mally esul ing in o e -enginee ing. Recen ad ances in in elligen compu ing b ings inspi a ion o new pe spec i es in he way we design, build, ope a ed and decommission ou sys ems. An enhanced s a egy is o ake a ious unce ain ies well in o accoun and, a he same ime, eplace High- Fideli y models wi h da a-suppo ed simple models (su oga e models) ha can be combined wi h Unce ain y Quan i ica ion (UQ). The e o e, he join use o su oga e models and UQ me hods o e s a po en ial solu ion, as i add esses enginee ing p oblems in a cos -e ec i e and echnically iable manne . The c ucial poin is o ensu e a comp ehensi e and accu a e UQ. Addi ionally, he no el oppo uni ies ha A i icial In elligence seem o o e in he domain pose speci ic challenges: (1) Ensu ing accu a e da a o AI/ML echniques; (2) Es ima ing AI/ML echnique p edic ion unce ain ies; (3) Explo ing AI/ML compliance wi h s anda ds and egula ions. 17 h In e na ional Con e ence on P obabilis ic Sa e y Assessmen and Managemen & Asian Symposium on Risk Assessmen and Managemen (PSAM17&ASRAM2024) 7-11 Oc obe , 2024, Sendai In e na ional Cen e , Sendai, Miyagi, Japan This pape is o ganised as ollows: a b ie o e iew o he ea men o unce ain y in nuclea sa e y isk assessmen s is p esen ed in Sec ion 2, ou lining he incen i es o he eMEANSS p ojec which aims o enhance he exis ing sa e y analysis and op imisa ion me hodologies. I includes an exposi ion o he di e en ypes o unce ain y con ibu ing o nuclea sa e y and he a ious sou ces om which hey a ise, he challenges in unce ain y modelling o a highly complex sys em such as a nuclea eac o . Sec ion 3 summa ises he s a e-o - he a compu a ional echniques pu ing speci ic ocus on us ul p edic i e modelling, imp ecise p obabili y and digi al win. 1.1. Enhanced Me hodologies o Ad anced Nuclea Sys em Sa e y (eMEANSS) p ojec Unce ain ies a e una oidable and complex sys ems such as nuclea eac o s a e designed o cope wi h hem. Imp ope app oaches, say conside ing indi idual wo s cases scena ios wi hou dependencies, would likely p oduce o e -designed and expensi e sys ems (i.e. conse a ism) wi hou gua an eeing hei o e all sa e y. By con as , p ope quan i ica ion and p opaga ion o unce ain y ac oss mul i-physical componen s allows one o de e mine ulne able componen y, p io i ise in es men s, iden i y ope a ional ma gins and adop ele an measu es o gua an ee sa e y whils a he same ime educing he o e all cos o ad anced nuclea design. The e o e, a e-assessmen o he impac o unce ain ies wi hin he nuclea indus y is o pa amoun impo ance, no only ensu ing he con inued sa e y o nuclea ene gy sys ems, bu also o ensu e he economic iabili y o new nuclea powe plan design, build, ope a ion and decommissioning. 2. UNCERTAINTY IN NUCLEAR ENGINEERING Sa e y analyses a e conduc ed o ensu e ha he design and ope a ional con ols o a nuclea acili y p o ide assu ance ha he public, s a , and he en i onmen a e p o ec ed om all nuclea haza ds. Owing o insu icien knowledge and unde s anding, conse a isms a e in oduced h oughou he sa e y analyses (e.g. in accep ance c i e ia, assump ions o models, inpu condi ions), such ha he assu ance o adequa e p o ec ion can be p o ided, suppo ing disciplines (e.g., quali y assu ance) and design p o isions (e.g. in- co po a ion o de ense-in-dep h and app op ia e sa e y ma gins) and ope a ional con ols. Nuclea powe echnology has been de eloped based la gely on he adi ional de ence-in-dep h philosophy o he design o he plan ha was suppo ed by de e minis ic and o e ly conse a i e me hods o sa e y analysis. In he pas , la ge unce ain ies in he compu e models used o nuclea powe sys em design and licensing ha e been compensa ed using highly conse a i e assump ions. Acciden scena ios a e ypically assessed using wo s case scena ios. Bes es ima e plus unce ain y (BEPU) is he leading me hodology in alida ing exis ing sa e y ma gins, bu i emains a challenge o de elop and license such app oaches. 2.1. P og ess on nuclea sa e y assessmen His o ical p og ess o he licensing app oach ha e gone h ough a ew phases: (I) Highly Conse a i e. (II) Realis ic Conse a i e; and (III) Use o Bes -Es ima e Plus Unce ain y (BEPU) [1]. Ini ially conse a i e hypo heses we e in oduced o sa e y analyses o add ess exis ing unce ain ies. Con en ional enginee ing is using sa e y ma gins in design in o de o compensa e o he unce ain y in modelling and simula ion, which ypically assumes wo s -case scena ios, in en ional o e es ima ion o pa ame e s, and esul s in o e design. This de e minis ic me hod, besides being cos ly, p o ides no way o es ima e isk o de e mine ailu e p obabili y and, hus, equi es he use o heu is ic sa e y ac o in an a emp o a oid ailu es. Wi h highly conse a i e assump ions, la ge unce ain ies in he compu e models used o nuclea powe sys em design ha e been compensa ed. The Loss-O -Coolan -Acciden E alua ion Model is one o he main examples abou his app oach. The use o mul iple conse a i e hypo hesis can in la e o ex emely conse a i e esul s bu i is claimed ha a easonable deg ee o conse a ism mus be sough in nuclea sa e y analyses o s ike he balance be ween sa e y and cos . In he absence o ull p opaga ion o pa ame e unce ain ies, he use o mean alues is consis en wi h he easonable conse a ism [2]. 17 h In e na ional Con e ence on P obabilis ic Sa e y Assessmen and Managemen & Asian Symposium on Risk Assessmen and Managemen (PSAM17&ASRAM2024) 7-11 Oc obe , 2024, Sendai In e na ional Cen e , Sendai, Miyagi, Japan BEPU equi es eplacing subjec i e judgmen s abou he inadequacy o he deg ee o conse a ism in he assump ions wi h quan i a i e measu es, which en ails he p opaga ion o code inpu unce ain y (selec ed numbe o pa ame e s) h ough he code o ob ain he ou pu unce ain y (e.g. p obabili y dis ibu ion unc ion) ei he ia code-nodalisa ion o epea ed code uns [3, 4, 5]. A concep ual compa ison, in e ms o p os and cons, be ween he conse a i e and BEPU app oaches can be ound in [1]. Public conce ns o nuclea sa e y and wide acknowledge o unce ain ies leads o p obabilis ic amewo ks in he managemen o unce ain y and isks o egula o y decision in he sa e y assessmen . P obabilis ic isk assessmen (PRA), as se ou in he Rasmussen Repo [6], in ol es de ining a sys em ailu e o complex mul icomponen , mul iphysics p oblems, iden i ying basic e en s ha can cause a sys em ailu e, building a aul ee o ela e componen e en s o a sys em ailu e, and ela ing he join p obabili y o e en s o he p obabili y o a sys em ailu e. I usually has a goal o mi iga e he isk o nuclea sa e y decisions in he p esence o unce ain y, h ough a comp ehensi e amewo k o calib a ion and he agg ega ion o isk in o ma ion om a ious sou ces, including nume ical calcula ions, expe opinions, isk a i udes, egula o y compliance, e c. [7, 8]. 2.2. Challenges and ecen de elopmen s on he cha ac e isa ion and quan i ica ion o unce ain y The e p esen s many challenges in a emp ing o app op i ely modelling he unce ain y, such as he unde s anding and modelling o he complica ed physics in e ms o a coupled mul iphysical and nonlinea sys em ha is nume ically ha d o sol e, he inc easing needs o powe p oduc ion cons ained by he ini ial sa e y design limi s on he basis o gene ally con en ional conse a ism ools, e c. No ably, a subs an ial challenge associa ed wi h unce ain y quan i ica ion o nuclea eac o designs is he necessi y o p opaga - ing unce ain ies h ough se e al linked simula ion codes o all o he coupled subsys ems. Hea ans e , coolan low, neu on dis ibu ions, and ission eac ion a es a e all igh ly coupled o o m a highly com- plex mul icompound, mul iphysics sys em. The esul ing models and simula ion codes a e compu a ionally in ensi e pe se, and i is compounded by he needs o comp ehensi e cha ac e isa ion o unce ain y along he pipeline o con iden ly desc ibing he quan i y o in e es (QoI). Ne e heless, UQ emains an ac i e a ea o esea ch in eac o physics/analysis and many o he sub- ields. [9] applied To al Mon e Ca lo me hodology o nuclea da a unce ain y p opaga ion o usion neu onics calcula ions and a numbe o usion shielding benchma ks. A e iew o UQ applica ion o compu a ional luid dynamics (CFD) analyses o nuclea eac o he mal hyd aulics can be ound in [10] and [11]. [12] in es iga es he measu emen unce ain y based on he powe moni o ing da a o a MIT esea ch eac o (MITR-II). [13] p oposes a new me hod ha cha ac e ises imp ecise and ague knowledge o de ec ing abno mal componen o he sys em unde unce ain y om he ins umen and con ol sys em o nuclea plan sys em. 2.3. Sou ce o unce ain y Unce ain ies a ise in many aspec s o nuclea eac o sys em modelling: in he nuclea da a, in he co e geome y, in he simula ion me hods, and in he plan da a wi h which simula ion esul s a e compa ed, e c. Gene ally hose can be ca ego ised in o se e al sou ces o unce ain ies and e o s: inpu pa ame e unce ain ies, model e o s, nume ical e o s, and da a unce ain ies (measu emen imp ecision, spa se and e en incomple e obse a ions) [9, 14, 15]. Pa icula ly, in de eloping ools and p ocedu es o nuclea eme gencies, [16] iden i ies 9 ypes o unce ain y: alea o ic, epis emic, ac o , judgemen al, compu a ional, model unce ain y as well unce ain ies ela ed o ambigui y and lack o cla i y, alue, social and e hical aspec s, and inally unce ain y abou he dep h o modelling. In p ac ice, limi ed plan da a is a ailable o bo h alida ing compu a ional models and de e mining he ela i e con ibu ions o o e all unce ain ies in such obse a ions. Nuclea da a e alua ions would 17 h In e na ional Con e ence on P obabilis ic Sa e y Assessmen and Managemen & Asian Symposium on Risk Assessmen and Managemen (PSAM17&ASRAM2024) 7-11 Oc obe , 2024, Sendai In e na ional Cen e , Sendai, Miyagi, Japan s a is ically mix expe imen al eac ion da a wi h eac ion models o p oduce he bes es ima e o he nuclea da a quan i ies plus unce ain y. Howe e , due o he di icul y and cos o conduc ing nuclea eac ion expe imen s, expe imen al da a is o en spa se o no p esen o he as majo i y o nuclides and eac ions. Excluding he main ission ela ed nuclides, which ha e been ex ensi ely s udied and ha e low unce ain ies, he unce ain y may be se e e o mos nuclides, wi h he wo s case being ha he co a iance in o ma ion is comple ely missing in some e alua ions [17]. The ision o a obus sa e y assessmen wo k low s a s om a sys ema ic app oach owa ds nuclea da a e alua ion, in which unce ain ies a e aken in o accoun in an app op ia e way, and which elies on e icien high- ideli y nuclea eac ion models ( o be de eloped) and high-p ecision measu emen s ( o be pe o med) [9]. The cu en impe ec ions wi h espec o he aspec s abo e (nuclea physics expe imen s, models and pa ame e s) d i e he needs o app op ia ely accoun o he a ious sou ces o unce ain y o inc easing o e all con idence in nuclea sa e y. 2.4. Modelling unce ain y alea o y unce ain y add esses inhe en a iabili y in sys ems ha canno be elimina ed e en wi h com- ple e knowledge. Also known as i educible unce ain y, alea o y unce ain y is a ibu ed o inhe en andomness o a iabili y in na u al phenomena. epis emic unce ain y deals wi h unce ain ies a ising om incomple e knowledge o lack o in o ma ion abou a sys em. This ype o unce ain y is o en con- side ed educible h ough addi ional da a collec ion, esea ch, o imp o ed modelling echniques. While p obabili y heo y has been used as he o hodox ool o alea o ic unce ain y, he e a e mo e discussions and heo ies as o o mula ing epis emic unce ain y ia non-p obabilis ic app oaches [18, 19], which en ails using in e als [20] and uzzy a iables [21] and seeks o na ow down unce ain y anges by upda ing mod- els as new in o ma ion becomes a ailable. I plays a c ucial ole in e ining p edic ion anges and making in o med decisions by con inuously imp o ing ou unde s anding o unce ain ac o s. Pa icula ly, mixed unce ain y model has a ac ed signi ican a en ion in ecen yea s h ough he de elopmen s o gene alized p obabili y heo ies (i.e., imp ecise p obabili y, Demps e –Sha e heo y) [20, 22], whe e unce ain numbe s (wi h ypical examples such as p obabili y boxes and c edal se s) play a ounda ional ole in many mode n isk assessmen s o complex enginee ing sys ems [23]. By compa ison, hie achical Bayesian me hods ( he second-o de dis ibu ion) also ep esen mixed ype o unce ain y bu uses p obabili y dis ibu ion o ac- coun o he unce ain y on shape pa ame e s. Fig. 1 illus a i ely displays hese models (p imi i es) o ep esen ing unce ain y. σ2 (a) p obabili y dis ibu ion (gaussian) (b) p obabili y box µ (c) highe -o de dis ibu ion (Gaussian in e se gamma) P obabili y P obabili y 17 h In e na ional Con e ence on P obabilis ic Sa e y Assessmen and Managemen & Asian Symposium on Risk Assessmen and Managemen (PSAM17&ASRAM2024) 7-11 Oc obe , 2024, Sendai In e na ional Cen e , Sendai, Miyagi, Japan (d) p obabili y mass unc ion (ca ego i- cal) (e) c edal se s ( ) highe -o de dis ibu ion (Di ichle ) Figu e 1: An illus a ion o he ma hema ical objec s (p imi i es) ep esen ing unce ain y 2.5. P opaga e unce ain y Complex enginee ing sys ems such as nuclea eac o exhibi ich unce ain y om a ious componen s and hie a chies. Mode n isk analyses o complex sys ems ca e ully dis inguish alea o ic and epis emic unce ain y (also e e ed o as a iabili y and ince i ude). The e a e he e o e challenges in he unce ain y analysis, a a sys em le el, o ep esen , agg ega e, and p opaga e mixed unce ain y ypes. The Mon e- Ca lo (MC) simula ion may se e as one o he mos widely used me hods o p opaga e alea o y unce ain y. Wi h MC i is possible o ea he model unde s udy as a black box, enabling non-in usi e unce ain y p opaga ion o be pe o med in abundan esea ch and applica ion domains. Howe e , i is wo hy o men ion ha such e sa ili y gene ally comes a he cos o in ensi e compu a ions o a nonlinea and high-dimensional complex sys em due o i s b u e- o ce na u e [24]. Wi h his said, he p og ess o e icien MC a ian s should also be well ecognised [25]. Al e na i ely, in e al p opaga ion is appealing as i can igo ously cap u e he unce ain y in QoI by yielding bounds ia in e al a i hme ic. Bu i s applicali y p esen s a challenge due o he lack o open- sou ce code nume ical codes in simula ions. When a mix u e o alea o y and epis emic unce ain y is p esen in he inpu , cha ac e ised by second-o de dis ibu ions o p-boxes, a numbe o di e en modi ica ions ha e been p oposed, including double-loop Mon e-Ca lo, gene alized impo ance sampling and in e al Mon e Ca lo [26]. One o he key challenges in deli e ing isk-based design op imisa ion o a mul iphysics sys em is he compu a ional bu den h ough high- ideli y nume ical simula ions, coupled by he conside a ion o unce - ain y. An e icien way o alle ia e such bu den is su oga e models which se e as e icien subs i u es, cap u ing he unde lying ela ionships be ween inpu and ou pu a iables in a non-in usi e ashion. By building a su oga e model, which is a simpli ied ma hema ical ep esen a ion o he o iginal model, one can signi ican ly educe compu a ional cos s while main aining a easonable le el o accu acy. Su oga e Models a e pa icula ly use ul o op imisa ion, sensi i i y analysis, and unce ain y quan i ica ion asks. In consid- e ing mixed alea o y-epis emic unce ain y quan i ica ion, which is highly challenging o complex nume ical codes, [27] p oposed o p opaga e p obabili y boxes h ough in e al p edic o models. [28] deals wi h he p opaga ion o unce ain y in he inpu pa ame e s cha ac e ised as p obabili y boxes h ough a de e minis ic, black-box compu a ional model. Fu he mo e, wo non-in usi e unce ain y p opaga ion app oaches a e p oposed in [29] o he analysis o gene ic enginee ing sys ems subjec o in e al unce ain ies. 3. TOWARDS AUTONOMOUS DIGITAL TWINS FOR NUCLEAR SYSTEMS: UNCERTAINTY, DATA AND AI Recen ad ancemen s in digi alisa ion and AI analy ics o e no el and p omising pe spec i es o mod- e nised app oaches o challenges and isions in nuclea unce ain y quan i ica ion and design op imisa ion. No ably, Digi al Twins (DT) o e he possibili ies o connec ing he i ual and physical wo lds o o e see he pe o mance o an asse , iden i y po en ial aul s and suppo be e -in o med decisions. Nuclea DT has been buil o accele a e he de elopmen and deploymen o ad anced nuclea echnology in a eas o passi e sa e y, new uel o ms, ins umen a ion, and eac o con ol [30], demons a ing po en ials in applica ions o p edic i e main enance, au onomous nuclea eac o con ol sys em, concep ual design op imisa ion, and imp o ed p ojec managemen , nuclea uel manu ac u ing, e c [31]. Wi h he buil physical asse , he DT collec eal- P obabili y 17 h In e na ional Con e ence on P obabilis ic Sa e y Assessmen and Managemen & Asian Symposium on Risk Assessmen and Managemen (PSAM17&ASRAM2024) 7-11 Oc obe , 2024, Sendai In e na ional Cen e , Sendai, Miyagi, Japan ime da a, e.g. om sma senso s, o unde s and he s a us, moni o he heal h o he sys em, p edic u u e scena ios and imp o ing he ideli y o he simula ion by dynamically upda ing he DT based on e idence. DT ep esen s an in eg a ed amewo k o calib a ion, da a assimila ion, unce ain y-in o med decision- making, planning and con ol, h ough, o ins ance, a p obabilis ic g aphical model [32]. I dynamically upda ed asse -speci ic compu a ional models in eg a ed wi hin he da a-d i en analysis and decision-making eedback loop [32]. The digi al win acqui es and assimila es obse a ional da a om he asse (e.g., da a om senso s o manual inspec ions) and uses his in o ma ion o con inually upda e i s in e nal models, e.g. Deep Lea ning (DL) models, so ha hey e lec he e ol ing physical sys em, which embodies a syne gis ic mul i-way coupling be ween he physical sys em, he da a collec ion, he compu a ional models, and he decision-making p ocess. One o he signi ican challenges is he unce ain y quan i ica ion [33, 34] which anges om o e o s in machine-lea ning models and low-quali y senso s, unce ain ies in oduced by simula ions, da a, and machine lea ning su oga e models. 3.1. P edic i e modelling wi h unce ain y awa eness The e ec i eness o AI (e.g. DL) analy ics ha e been widely ecognised and u ilised in he nuclea powe indus y chain o ele a e da a analyses and decision making a a ious s eps such as nuclea uel supply (ups eam), nuclea equipmen manu ac u ing (mids eam), and he nuclea powe plan design, ope a ion, and main enance (downs eam) [35, 36, 37, 38]. Howe e , da a could be impe ec as being spa se, sca ce, and imp ecise. Lea ning om da a o insu icien quali y has es ic ed he e ec i eness o da a-d i en echniques in lea ning he ue unde lying da a gene a ing p ocess, which u he deg ades he pe o mance o gene alisa ion. I is equi ed in many sa e y-c i ical applica ions o consequen ial enginee ing p ac ices ha he model should be capable o signalling when i is unce ain o i s esul s (i.e. know when hey do no know) o be obus and us ul, as opposed o o e - con iden ly yielding inaccu a e p edic ions/ o ecas s/decisions. Inc easing a en ion has been he e o e ocused on he de elopmen s o T us wo hy AI which aims o c i ically in es iga e he ai ness (biasness), in e p e abili y, and obus ness o Deep Lea ning algo i hms and applica ions. Reso ing o p io knowledge is an e ec i e app oach agains da a insu iciency, [39] p oposes a me a-lea ning app oach o obus ly p edic ma e ial p ope ies o nuclea eac o design unde limi ed da a. Imp ecise p obabili y amewo k u he ele a es he capaci y o machine lea ning models in accoun ing o unce ain y, especially when dealing wi h pa ial o ague knowledge, such as when he a ailable da a is ai ly limi ed such ha a p ecise speci ica ion o p obabili y densi y unc ion (PDF) canno be de e mined wi h con idence [40, 41, 42], o dealing wi h da a ince i ude (e en in ex eme cases as missing da a) [43, 44, 45]. Impo an ly, IP enables he conside a ion o unce ain y in decision-making based on he p edic i e ou comes as i ealis ically cha ac e ises indecision whe e he cu en e idence is inadequa e o yield a decision wi hou o e con idence. Using AI analy ics can also be bene icial in enhancing esilience and sa e y. Con a y o a human ope a o , AI sys ems could analyse huge amoun o da a and p edic he consequences o he decision in c i ical si ua ion wi hou su e ing om ypical human ela ed e o s due o s ess and en i onmen al and o ganisa ional p essu e [46]. Possessing he po en ial o enable au onomous con ol sys em (e.g. in nuclea powe plan s) [47], e en as no implemen ed as ully au onomous, DT and AI can be used o pass only he mos ele an in o ma ion wi h clea le el o “c edibili y” o a decision make . AI can be employed o p edic ing po en ial mal unc ions and au onomously making p oac i e decisions. 3.2. Tools and so wa e coupling unce ain y quan i ica ion and deep lea ning COSSAN a gene alised so wa e o unce ain y quan i ica ion in isk, eliabili y and esilience analy- ses [48, 49]. P obabilis ic P og amming Languages (PPL) ha empowe ing a wide spec um o Bayesian machine lea ning me hods [50, 51, 52] 17 h In e na ional Con e ence on P obabilis ic Sa e y Assessmen and Managemen & Asian Symposium on Risk Assessmen and Managemen (PSAM17&ASRAM2024) 7-11 Oc obe , 2024, Sendai In e na ional Cen e , Sendai, Miyagi, Japan Figu e 2. An o e iew o he p og ams and so wa e coupling unce ain y quan i ica ion wi h deep lea ning 4. CONCLUSION I has been widely ecognised he needs o enhanced sa e y analyses and op imisa ion o nuclea sys ems o imp o e upon conse a ism. Con en ional conse a ism canno es ablish he sa e y ma gins in a quan i a i e manne nei he achie e he op imisa ion o he sa e y solu ion. This pape se s ou an enhanced s a egy ha le e ages he s a e-o - he-a compu ing echnologies, aking a ious unce ain ies well in o accoun , meanwhile, eplacing High-Fideli y models wi h da a-suppo ed su oga e models (e.g. us ul Machine Lea ning (ML) models). This new s a egy embodies a mu ual coupling be ween he physical sys em, da a collec ion, compu a ional models, and he decision-making p ocess. No ably, digi al win, as an eme ging echnology, se es as an in eg a ed hub o calib a ion, da a assimila ion, unce ain y-in o med decision- making, planning and con ol. Imp ecise p obabili y p o ides a igo ous ep esen a ions o a ious unce ain ies, including a iabili y, imp ecision, and agueness, o be combined and exp essed by he uni ied ma hema ical s uc u e. The combina ion o ML and IP he e o e leads o an ele a ed le el o c edibili y whe e eliable decision can be based upon. By connec ing he i ual and physical wo lds, Digi al wins o e he possibili ies o o obus ly o e see he pe o mance o an asse , iden i y po en ial aul s and suppo be e - in o med decisions. The enhanced s a egy, coupled wi h he expe ise om mul idisciplina y and mul i physical pe spec i es, has huge po en ials o add ess c i ical challenges in e i ica ion, alida ion, and unce ain y quan i ica ion o he sa e and con iden use o digi al win echnologies in ad anced eac o s. Acknowledgemen s This wo k was suppo ed Enhanced Me hodologies o Ad anced Nuclea Sys em Sa e y (eMEANSS) [P ojec no. EP/T016329/1]. Re e ences [1] A. Bucalossi, e al., Cu en use o bes es ima e plus unce ain y me hods on ope a ional p ocedu es add essing no mal and eme gency condi ions (2008). [2] K. Jamali, Achie ing easonable conse a ism in nuclea sa e y analyses, Reliabili y Enginee ing & Sys em Sa e y 137 (2015) 112–119. [3] F. D’Au ia, Bes es ima e plus unce ain y (bepu): s a us and pe spec i es, Nuclea Enginee ing and Design 352 (2019) 110190. [4] F. S. D’Au ia, H. Glaese , S. Lee, J. Mi´ak, M. Mod o, R. Schul z, e al., Bes es ima e sa e y analysis o nuclea powe plan s: Unce ain y e alua ion. iaea sa e y epo se ies (2008). [5] R. P. Ma in, A. Pe uzzi, P og ess in in e na ional bes es ima e plus unce ain y analysis me hodologies, Nuclea Engi- nee ing and Design 374 (2021) 111033. [6] C. Gilbe , Nuclea eac o sa e y-a e iew o he Rasmussen Repo (WASH-1400)., Aus alian A omic Ene gy Commis- sion, 1979. [7] R. C. Smi h, Unce ain y quan i ica ion: heo y, implemen a ion, and applica ions, SIAM, 2013. [8] M. Pa ´e-Co nell, P obabili y and unce ain y in nuclea sa e y decisions, Nuclea Enginee ing and Design 93 (2-3) (1986) 319–327. 17 h In e na ional Con e ence on P obabilis ic Sa e y Assessmen and Managemen & Asian Symposium on Risk Assessmen and Managemen (PSAM17&ASRAM2024) 7-11 Oc obe , 2024, Sendai In e na ional Cen e , Sendai, Miyagi, Japan [9] D. Rochman, A. Koning, S. Van De Ma ck, Exac nuclea da a unce ain y p opaga ion o usion neu onics calcula ions, Fusion Enginee ing and Design 85 (5) (2010) 669–682. [10] D. Bes ion, A. De C ecy, F. Mo e i, R. Camy, A. Ba he , S. Belle , J. M. Cobo, A. Badillo, B. Niceno, P. Hedbe g, e al., Re iew o unce ain y me hods o c d applica ion o nuclea eac o he malhyd aulics, in: NUTHOS 11-The 11 h In e na ional Topical Mee ing on Nuclea Reac o The mal Hyd aulics, Ope a ion and Sa e y, 2016. [11] M. Pou gol-Mohammad, The mal–hyd aulics sys em codes unce ain y assessmen : A e iew o he me hodologies, Annals o Nuclea Ene gy 36 (11-12) (2009) 1774–1786. [12] R. R. E. Pa us, Quan i ying measu emen unce ain ies o mi esea ch eac o ’s new digi al nuclea sa e y sys em, Ph.D. hesis, Massachuse s Ins i u e o Technology (2018). [13] R. kamal Kau , L. K. Singh, A. Khampa ia, Modeling unce ain y o ins umen and con ol sys em o nuclea powe plan , Annals o Nuclea Ene gy 139 (2020) 107207. [14] Y. Chen, E. Pa elli, B. Edwa ds, M. Bee , A bayesian augmen ed-lea ning amewo k o spec al unce ain y quan i ica ion o incomple e eco ds o s ochas ic p ocesses, Mechanical Sys ems and Signal P ocessing 200 (2023) 110573. [15] D. Smi h, Nuclea da a unce ain y quan i ica ion: pas , p esen and u u e, Nuclea Da a Shee s 123 (2015) 1–7. [16] S. F ench, S. Haywood, D. Ough on, C. Tu canu, Di e en ypes o unce ain y in nuclea eme gency managemen , Radiop o ec ion 55 (suppl. 1) (2020) S175–S180. [17] A. G ay, A. Da is, E. Pa elli, Unce ain y p opaga ion in sinbad usion benchma ks wi h o al mon e ca lo and imp ecise p obabili ies, Fusion Science and Technology 77 (7-8) (2021) 802–812. [18] M. Faes, D. Moens, Recen ends in he modeling and quan i ica ion o non-p obabilis ic unce ain y, A chi es o Com- pu a ional Me hods in Enginee ing 27 (2020) 633–671. [19] A. G ay, S. Fe son, V. K eino ich, E. Pa elli, Dis ibu ion- ee isk analysis, In e na ional Jou nal o App oxima e Rea- soning 146 (2022) 133–156. [20] S. Fe son, J. G. Hajagos, A i hme ic wi h unce ain numbe s: igo ous and (o en) bes possible answe s, Reliabili y Enginee ing & Sys em Sa e y 85 (1-3) (2004) 135–152. [21] B. M¨olle , M. Bee , Fuzzy andomness: unce ain y in ci il enginee ing and compu a ional mechanics, Sp inge Science & Business Media, 2013. [22] S. Fe son, W. L. Obe kamp , Valida ion o imp ecise p obabili y models, In e na ional Jou nal o Reliabili y and Sa e y 3 (1-3) (2009) 3–22. [23] M. Bee , S. Fe son, V. K eino ich, Imp ecise p obabili ies in enginee ing analyses, Mechanical sys ems and signal p ocessing 37 (1-2) (2013) 4–29. [24] P. Zeng, T. Li, Y. Chen, R. Jimenez, X. Feng, S. Senen , New colloca ion me hod o s ochas ic esponse su ace eliabili y analyses, Enginee ing wi h Compu e s 36 (2020) 1751–1762. [25] M. de Angelis, E. Pa elli, M. Bee , Ad anced line sampling o e icien obus eliabili y analysis, S uc u al sa e y 52 (2015) 170–182. [26] A. G ay, A. Wimbush, M. de Angelis, P. O. H is o , D. Calleja, E. Mi alles-Dolz, R. Rocche a, F om in e ence o design: A comp ehensi e amewo k o unce ain y quan i ica ion in enginee ing wi h limi ed in o ma ion, Mechanical Sys ems and Signal P ocessing 165 (2022) 108210. [27] J. Sadeghi, M. De Angelis, E. Pa elli, Robus p opaga ion o p obabili y boxes by in e al p edic o models, S uc u al Sa e y 82 (2020) 101889. [28] R. Sch¨obi, B. Sud e , Unce ain y p opaga ion o p-boxes using spa se polynomial chaos expansions, Jou nal o Compu- a ional Physics 339 (2017) 307–327. [29] A. Cici ello, F. Giun a, Machine lea ning based op imiza ion o in e al unce ain y p opaga ion, Mechanical Sys ems and Signal P ocessing 170 (2022) 108619. [30] N. C espi, A. T. D obo , R. Mine a, The digi al win: Wha and why?, in: The Digi al Twin, Sp inge , 2023, pp. 3–20. [31] M. Bandala, P. Cha d, N. Cockbain, D. Dunphy, D. Ea es, D. Hu chinson, D. Lee, X. Ma, S. Ma shall, P. Mu ay, e al., Digi al win challenges and oppo uni ies o nuclea uel manu ac u ing applica ions, Nuclea Enginee ing and Design 420 (2024) 113013. [32] M. G. Kap eyn, J. V. P e o ius, K. E. Willcox, A p obabilis ic g aphical model ounda ion o enabling p edic i e digi al wins a scale, Na u e Compu a ional Science 1 (5) (2021) 337–347. [33] V. Yada , V. Aga wal, A. V. G ibok, R. D. Hays, A. J. Plu h, C. S. Ri e , H. Zhang, P. K. Jain, P. Ramuhalli, D. Eskins, e al., Technical challenges and gaps in digi al- win-enabling echnologies o nuclea eac o applica ions, US Nuclea Regula o y Commission, Washing on, DC (2021). [34] B. Kochunas, X. Huan, Digi al win concep s wi h unce ain y o nuclea powe applica ions, Ene gies 14 (14) (2021) 4235. [35] C. Tang, C. Yu, Y. Gao, J. Chen, J. Yang, J. Lang, C. Liu, L. Zhong, Z. He, J. L , Deep lea ning in nuclea indus y: A su ey, Big Da a Mining and Analy ics 5 (2) (2022) 140–160. [36] R. P. Ma in, B. Nasse sha i , Deep knowledge expe sys em o diagnosis o mul iple- ailu e se e e ansien s in nuclea powe plan , in: A i icial In elligence and O he Inno a i e Compu e Applica ions in he Nuclea Indus y, Sp inge , 1988, pp. 281–288. [37] D. Lee, J. Kim, Au onomous algo i hm o sa e y sys ems o he nuclea powe plan by using he deep lea ning, in: Ad ances in Human Fac o s in Ene gy: Oil, Gas, Nuclea and Elec ic Powe Indus ies: P oceedings o he AHFE 2017 In e na ional Con e ence on Human Fac o s in Ene gy: Oil, Gas, Nuclea 17 h In e na ional Con e ence on P obabilis ic Sa e y Assessmen and Managemen & Asian Symposium on Risk Assessmen and Managemen (PSAM17&ASRAM2024) 7-11 Oc obe , 2024, Sendai In e na ional Cen e , Sendai, Miyagi, Japan and Elec ic Powe Indus ies, July 17–21, 2017, The Wes in Bona en u e Ho el, Los Angeles, Cali o nia, USA 8, Sp inge , 2018, pp. 72–82. [38] O. Toka li, P. Das, R. Na h, L. Pangione, A. Al obelli, G. Bu oughes, E. T. Jonasson, M. F. Tu ne , R. Skil on, Robo - assis ed glo ebox eleope a ion o nuclea indus y, Robo ics 10 (3) (2021) 85. [39] Y. Chen, E. Pa elli, Y. Zhen, A. Lye, Towa ds obus p edic ion o ma e ial p ope ies o nuclea eac o design unde sca ce da a – a s udy in c eep up u e p ope y, a Xi p ep in a Xi :2405.17862 (2024). [40] F. G. Cozman, Lea ning imp ecise p obabili y models: Concep ual and p ac ical challenges, In e na ional Jou nal o App oxima e Reasoning 55 (7) (2014) 1594–1596. [41] E. Quaeghebeu , G. De Cooman, Imp ecise p obabili y models o in e ence in exponen ial amilies, in: 4 h In e na ional Symposium on Imp ecise P obabili ies and Thei Applica ions, In e na ional Socie y o Imp ecise P obabili y: Theo ies and Applica ions . . . , 2005, pp. 287–296. [42] G. Wal e , T. T. Augus in, A. Pe e s, Linea eg ession analysis unde se s o conjuga e p io s, in: Fi h In e na ional Sym- posium on Imp ecise P obabili ies and hei Applica ions, Socie y o Imp ecise P obabili y: Theo ies and Applica ions, 2007, pp. 445–455. [43] T. Augus in, R. Hable, On he impac o obus s a is ics on imp ecise p obabili y models: a e iew, S uc u al Sa e y View publica ion s a s